text
stringlengths
62
4.5k
arxiv-translator
stringlengths
51
2.86k
In Section 4 we showed that we can express the (unavailable) ground-truth reward through its corresponding optimal policy:
์„น์…˜ 4์—์„œ ์šฐ๋ฆฌ๋Š” ์ƒ์‘ํ•˜๋Š” ์ตœ์  ์ •์ฑ…์„ ํ†ตํ•ด (์ด์šฉํ•  ์ˆ˜ ์—†๋Š”) ์ง€์ƒ ์ง„์‹ค ๋ณด์ƒ์„ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค:
Most commercial LLMs, like ChatGPT (OpenAI, 2023), Gemini (Gemini Team, 2023), and Perplexity.ai already employ some version of RAG in their Web interfaces. For example, ChatGPT employs a Bing search whereas Gemini accesses Google Search results. Though RAG has quickly become a default feature of user-facing LLM systems, most evaluations of LLM capabilities are still performed on the non-RAG counterparts (Zheng et al., 2023). This is problematic, as a model's default and RAG-enabled responses can drastically diverge depending on the quality and accuracy of the retrieved content. This problem is compounded when considering that web results constantly change, and can contain outdated, incorrect, or harmful information (Dash et al., 2023; Daws, 2020; Nastasi et al., 2023). Thus, objective evaluations of RAG-enabled LLM behavior are as important as benchmarking their non-RAG counterparts, especially as RAG systems are increasingly relied upon to provide factual information in a myriad of domains.
ChatGPT(OpenAI, 2023), Gemini(Gemini Team, 2023), Perplexity.ai์™€ ๊ฐ™์€ ๋Œ€๋ถ€๋ถ„์˜ ์ƒ์šฉ LLM์€ ์ด๋ฏธ ์›น ์ธํ„ฐํŽ˜์ด์Šค์— ์ผ๋ถ€ ๋ฒ„์ „์˜ RAG๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ChatGPT๋Š” Bing ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ˜๋ฉด Gemini๋Š” Google ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์•ก์„ธ์Šคํ•ฉ๋‹ˆ๋‹ค. RAG๊ฐ€ ์‚ฌ์šฉ์ž-๋Œ€๋ฉด LLM ์‹œ์Šคํ…œ๋“ค์˜ ๋””ํดํŠธ ํŠน์ง•์ด ๋น ๋ฅด๊ฒŒ ๋˜์—ˆ์ง€๋งŒ, LLM ๋Šฅ๋ ฅ๋“ค์˜ ๋Œ€๋ถ€๋ถ„์˜ ํ‰๊ฐ€๋“ค์€ ์—ฌ์ „ํžˆ ๋น„-RAG ๋Œ€์‘๋ฌผ๋“ค์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋œ๋‹ค(Zheng et al., 2023). ์ด๋Š” ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ๊ฐ’ ๋ฐ RAG ์ง€์› ์‘๋‹ต์ด ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ๊ณผ ์ •ํ™•๋„์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์ œ๊ฐ€ ๋œ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ์›น ๊ฒฐ๊ณผ๊ฐ€ ์ง€์†์ ์œผ๋กœ ๋ณ€ํ™”ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•  ๋•Œ ๋” ๋ณต์žกํ•ด์ง€๋ฉฐ, ๊ตฌ์‹, ๋ถ€์ •ํ™•ํ•œ ๋˜๋Š” ์œ ํ•ดํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ๋‹ค(Dash et al., 2023; Daws, 2020; Nastasi et al., 2023). ๋”ฐ๋ผ์„œ RAG ์ง€์› LLM ํ–‰๋™์— ๋Œ€ํ•œ ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋Š” ํŠนํžˆ RAG ์‹œ์Šคํ…œ์ด ๋ฌด์ˆ˜ํžˆ ๋งŽ์€ ๋„๋ฉ”์ธ์—์„œ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์ ์  ๋” ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋น„ RAG ๋Œ€์‘๋ฌผ์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๋Š” ๊ฒƒ๋งŒํผ ์ค‘์š”ํ•˜๋‹ค.
For the \(1000\)-exposure setting, the model's final performance is _not very sensitive_ to learning rate choices due to sufficient training. The following parameters were chosen for generating Figure 1(a):
\(1000\)-๋…ธ์ถœ ์„ค์ •์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ์˜ ์ตœ์ข… ์„ฑ๋Šฅ์€ ์ถฉ๋ถ„ํ•œ ํ›ˆ๋ จ์œผ๋กœ ์ธํ•ด ํ•™์Šต๋ฅ  ์„ ํƒ์— ๋งค์šฐ ๋ฏผ๊ฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ ํŒŒ๋ผ๋ฏธํ„ฐ๋“ค์ด ๋„ 1(a)๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์„ ํƒ๋˜์—ˆ๋‹ค:
Evaluating different LLMs can be seen also from different perspectives. For example, a LLM with a drastically fewer number of parameters is not completely comparable to one with a larger number of parameters. From this perspective, we will categorize LLMs in four categories as well: **small** (less than or equal to 1 billion parameters), **medium** (between 1 and 10 billion), **large** (between 10 and 100 billion), and **very large** (more than 100 billion). Another classification for the LLMs we use is their primary use case. We consider each LLM to be either: **Foundation** model (pretrained language model with no instruction fine-tuning and chat fine-tuning), **Instruction** model (pretrained language model with only instruction fine-tuning), and **Chat** model (pretrained language model with instruction and chat fine-tuning). Apart from all the categorization described, another category is required to distinguish between original models and tuned ones. **Original** models are those that have been released as a foundation model or a fine-tuned one. **Tuned** models are those that grasped the original model and tuned it with different datasets or even different training approaches. It is also good to note that original models are usually foundation models that have been fine-tuned on specific datasets or even different approaches. Availability of the model weights regardless of the license is another category in our classification. Models that have their weights publicly available (even through request) are noted as **Public** models while others are noted as **Private**. Table III shows all of these definitions and abbreviations used in the rest of the article. Figure 43 illustrate these visually.
๋‹ค๋ฅธ LLM์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋‹ค๋ฅธ ๊ด€์ ์—์„œ๋„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๊ธ‰๊ฒฉํžˆ ์ ์€ LLM์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๋” ๋งŽ์€ LLM๊ณผ ์™„์ „ํžˆ ๋น„๊ตํ•  ์ˆ˜ ์—†๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ ์—์„œ LLM์€ **์ž‘์€**(10์–ต ๋งค๊ฐœ ๋ณ€์ˆ˜ ์ดํ•˜), **์ค‘๊ฐ„**(10์–ต์—์„œ 100์–ต ์‚ฌ์ด), **ํฐ**(10์–ต์—์„œ 1000์–ต ์‚ฌ์ด) ๋ฐ **๋งค์šฐ ํฐ**(1000์–ต ์ด์ƒ)์˜ ๋„ค ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” LLMs์— ๋Œ€ํ•œ ๋˜ ๋‹ค๋ฅธ ๋ถ„๋ฅ˜๋Š” ๊ทธ๋“ค์˜ ์ฃผ์š” ์‚ฌ์šฉ ์‚ฌ๋ก€์ด๋‹ค. ๊ฐ LLM์€ **Foundation** ๋ชจ๋ธ(๋ช…๋ น ๋ฏธ์„ธ ์กฐ์ • ๋ฐ ์ฑ„ํŒ… ๋ฏธ์„ธ ์กฐ์ •์ด ์—†๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ), **๋ช…๋ น** ๋ชจ๋ธ(๋ช…๋ น ๋ฏธ์„ธ ์กฐ์ •๋งŒ ์žˆ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ) ๋ฐ **์ฑ„ํŒ…** ๋ชจ๋ธ(๋ช…๋ น ๋ฐ ์ฑ„ํŒ… ๋ฏธ์„ธ ์กฐ์ •์ด ์žˆ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ) ์ค‘ ํ•˜๋‚˜๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค. ์„ค๋ช…๋œ ๋ชจ๋“  ๋ฒ”์ฃผํ™” ์™ธ์—๋„ ์›๋ž˜ ๋ชจ๋ธ๊ณผ ์กฐ์ •๋œ ๋ชจ๋ธ์„ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค๋ฅธ ๋ฒ”์ฃผ๊ฐ€ ํ•„์š”ํ•˜๋‹ค. **์›๋ณธ** ๋ชจ๋ธ์€ ๊ธฐ์ดˆ ๋ชจ๋ธ ๋˜๋Š” ๋ฏธ์„ธ ์กฐ์ • ๋ชจ๋ธ๋กœ ์ถœ์‹œ๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. **ํŠœ๋‹** ๋ชจ๋ธ์€ ์›๋ž˜ ๋ชจ๋ธ์„ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ์‹ฌ์ง€์–ด ๋‹ค๋ฅธ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ํŠœ๋‹ํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์›๋ณธ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ์‹ฌ์ง€์–ด ๋‹ค๋ฅธ ์ ‘๊ทผ๋ฒ•์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •๋œ ๊ธฐ์ดˆ ๋ชจ๋ธ์ด๋ผ๋Š” ์ ์— ์ฃผ๋ชฉํ•˜๋Š” ๊ฒƒ๋„ ์ข‹๋‹ค. ๋ผ์ด์„ ์Šค์— ๊ด€๊ณ„์—†์ด ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์˜ ๊ฐ€์šฉ์„ฑ์€ ๋ถ„๋ฅ˜์˜ ๋˜ ๋‹ค๋ฅธ ๋ฒ”์ฃผ์ด๋‹ค. ์ค‘๋Ÿ‰์ด ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ(์š”์ฒญ์„ ํ†ตํ•ด์„œ๋„)์€ **๊ณต์šฉ** ๋ชจ๋ธ๋กœ ํ‘œ์‹œ๋˜๊ณ  ๋‹ค๋ฅธ ๋ชจ๋ธ์€ **๊ฐœ์ธ**์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ํ‘œ III์€ ๋‚˜๋จธ์ง€ ๊ธฐ์‚ฌ์—์„œ ์‚ฌ์šฉ๋œ ์ด๋Ÿฌํ•œ ์ •์˜ ๋ฐ ์•ฝ์–ด๋ฅผ ๋ชจ๋‘ ๋ณด์—ฌ์ค€๋‹ค. ๋„ 43์€ ์ด๋“ค์„ ์‹œ๊ฐ์ ์œผ๋กœ ๋„์‹œํ•œ ๊ฒƒ์ด๋‹ค.
* For the reversal tasks of determining the person's full name, only in the entity-preserving reversal case do accuracies become non-trivial. Both token/word reversals completely fail in such tasks. * When determining a person's full name given only the birth date, the reversal task accuracy remains near zero; this is because dates are treated as three entities in our adopted entity detection method, so their ordering is _not preserved_ in the reversal. * If the reversal tasks are simplified to determining the person's _last name only_, then word-level reversal suffices, and token-level reversal also yields non-trivial accuracies. * Some readers may find it surprising that an entity-preserving method can determine the person's full name but not the person's last name. This is a known phenomenon (Allen-Zhu & Li, 2023b): a language model may completely fail at retrieving _later_ tokens of a knowledge piece (such as the last name) without the help of spelling out earlier tokens -- they can be viewed as a Chain of Thought (CoT).
* ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ˜์ „ ์ž‘์—…์˜ ๊ฒฝ์šฐ ์—”ํ„ฐํ‹ฐ ๋ณด์กด ๋ฐ˜์ „ ๊ฒฝ์šฐ์—๋งŒ ์ •ํ™•์„ฑ์ด ์ค‘์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ† ํฐ/์›Œ๋“œ ๋ฐ˜์ „ ๋ชจ๋‘ ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ์™„์ „ํžˆ ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค. * ์ƒ๋…„์›”์ผ๋งŒ ์ฃผ์–ด์ง„ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•  ๋•Œ ๋ฐ˜์ „ ์ž‘์—… ์ •ํ™•๋„๋Š” 0์— ๊ฐ€๊น๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‚ ์งœ๊ฐ€ ์ฑ„ํƒ๋œ ์—”ํ„ฐํ‹ฐ ํƒ์ง€ ๋ฐฉ๋ฒ•์—์„œ ์„ธ ๊ฐœ์˜ ์—”ํ„ฐํ‹ฐ๋กœ ์ฒ˜๋ฆฌ๋˜๋ฏ€๋กœ ๋ฐ˜์ „์—์„œ ์ˆœ์„œ๊ฐ€ _๋ณด์กด๋˜์ง€_ ์•Š๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. * ๋ฐ˜์ „ ์ž‘์—…์„ ์‚ฌ์šฉ์ž์˜ _์„ฑ๋งŒ_ ๊ฒฐ์ •์œผ๋กœ ๋‹จ์ˆœํ™”ํ•˜๋Š” ๊ฒฝ์šฐ ๋‹จ์–ด ์ˆ˜์ค€ ๋ฐ˜์ „์œผ๋กœ ์ถฉ๋ถ„ํ•˜๊ณ  ํ† ํฐ ์ˆ˜์ค€ ๋ฐ˜์ „๋„ ์ค‘์š”ํ•˜์ง€ ์•Š์€ ์ •ํ™•๋„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. * ์ผ๋ถ€ ๋…์ž๋Š” ์—”ํ„ฐํ‹ฐ ๋ณด์กด ๋ฐฉ๋ฒ•์ด ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์‚ฌ๋žŒ์˜ ์„ฑ์€ ๊ฒฐ์ •ํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ด ๋†€๋ž๋‹ค๊ณ  ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ์•Œ๋ ค์ง„ ํ˜„์ƒ์ž…๋‹ˆ๋‹ค (Allen-Zhu & Li, 2023b): ์–ธ์–ด ๋ชจ๋ธ์€ ์ด์ „ ํ† ํฐ์„ ์ฒ ์žํ•˜์ง€ ์•Š๊ณ  ์ง€์‹ ์กฐ๊ฐ(์˜ˆ: ์„ฑ)์˜ _ํ›„๊ธฐ_ ํ† ํฐ์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ์™„์ „ํžˆ ์‹คํŒจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” CoT(์‚ฌ์Šฌ ์˜ค๋ธŒ ์‚ฌ์ƒ)๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* For \(N=50K\), we use \(wd=0.01\), \(lr=0.001\), and batch size 12; * For \(N=100K\), we use \(wd=0.01\), \(lr=0.001\), and batch size 24; * For \(N=200K\), we use \(wd=0.01\), \(lr=0.001\), and batch size 48; (except for GPT2-2-20, where \(lr=0.0005\) is used) * For \(N=500K\), we use \(wd=0.01\), \(lr=0.0005\), and batch size 96; * For \(N=1M\), we use \(wd=0.01\), \(lr=0.0005\), and batch size 192; * For \(N=2M\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 384; * For \(N=5M\), we use \(wd=0.01\), \(lr=0.0003/0.0005\), and batch size 768; * For \(N=10M\), we use \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\), and batch size 1024; * For \(N=20M\), we use \(wd=0.002\), \(lr=0.0002/0.0003/0.0005\), and batch size 1536.22
* \(N=50K\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.001\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 12๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=100K\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.001\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 24๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=200K\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.001\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 48์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. (GPT2-2-20์„ ์ œ์™ธํ•˜๊ณ , ์—ฌ๊ธฐ์„œ \(lr=0.0005\)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.) * \(N=500K\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.0005\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 96์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=1M\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.0005\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 192๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=2M\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.0003/0.0005/0.001\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 384๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=5M\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.0003/0.0005\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 768์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=10M\)์˜ ๊ฒฝ์šฐ \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 1024๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. * \(N=20M\)์˜ ๊ฒฝ์šฐ \(wd=0.002\), \(lr=0.0002/0.0003/0.0005\) ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 1536.22๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Mathematically, let \(\mathcal{M}\) be a pre-trained model, and let \(\mathcal{Q}\) be a set of factual questions, as before. Now, assume we have a relevant auxiliary knowledge base \(\mathcal{B}_{\mathcal{Q}}\). Our objective is to discover a transformation, denoted as \(\mathcal{F}\), that, when applied, would enhance the knowledge about \(\mathcal{Q}\):
์ˆ˜ํ•™์ ์œผ๋กœ, \(\mathcal{M}\)๋Š” ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด๊ณ , \(\mathcal{Q}\)๋Š” ์ด์ „๊ณผ ๊ฐ™์ด ์‚ฌ์‹ค์  ์งˆ๋ฌธ์˜ ์ง‘ํ•ฉ์ด๋ผ๊ณ  ํ•˜์ž. ์ด์ œ, ์šฐ๋ฆฌ๊ฐ€ ๊ด€๋ จ ๋ณด์กฐ ์ง€์‹ ๋ฒ ์ด์Šค \(\mathcal{B}_{\mathcal{Q}}\)๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•˜์ž. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” \(\mathcal{F}\)๋กœ ํ‘œ๊ธฐ๋œ ๋ณ€ํ™˜์„ ๋ฐœ๊ฒฌํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ๊ทธ๊ฒƒ์ด ์ ์šฉ๋  ๋•Œ, \(\mathcal{Q}\)์— ๋Œ€ํ•œ ์ง€์‹์„ ํ–ฅ์ƒ์‹œํ‚ฌ ๊ฒƒ์ด๋‹ค:
the data-matched baseline, showing that in the data-bound case, reverse training even helps on standard tasks. In the compute-matched case, where the baseline has effectively access to more data, reversal training lags slightly behind for 5, 10 samples. Importantly, in both the data-matched and compute-matched case we see significant improvement in the reverse direction questions for reverse training compared to either baseline. This demonstrates that reverse training can be employed during the pre-training stage to make the model robust against the reversal curse.
๋ฐ์ดํ„ฐ ๋งค์นญ ๊ธฐ์ค€์„ ์€ ๋ฐ์ดํ„ฐ ๋ฐ”์ธ๋”ฉ์˜ ๊ฒฝ์šฐ ์—ญ๋ฐฉํ–ฅ ํ›ˆ๋ จ์ด ํ‘œ์ค€ ์ž‘์—…์—๋„ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค. ๊ธฐ์ค€์„ ์ด ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ์— ํšจ๊ณผ์ ์œผ๋กœ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ„์‚ฐ ์ผ์น˜์˜ ๊ฒฝ์šฐ ๋ฐ˜์ „ ํ›ˆ๋ จ์€ 5, 10๊ฐœ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ์•ฝ๊ฐ„ ๋’ค์ฒ˜์ง„๋‹ค. ์ค‘์š”ํ•˜๊ฒŒ๋„, ๋ฐ์ดํ„ฐ ๋งค์นญ ๋ฐ ์ปดํ“จํŒ… ๋งค์นญ ์‚ฌ๋ก€ ๋ชจ๋‘์—์„œ ์šฐ๋ฆฌ๋Š” ์–ด๋Š ๊ธฐ์ค€์„ ์— ๋น„ํ•ด ์—ญ๋ฐฉํ–ฅ ํŠธ๋ ˆ์ด๋‹์— ๋Œ€ํ•œ ์—ญ๋ฐฉํ–ฅ ์งˆ๋ฌธ์—์„œ ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋ณธ๋‹ค. ์ด๋Š” ์—ญ์ „ ์ €์ฃผ์— ๊ฐ•๊ฑดํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ ์—ญ์ „ ํ›ˆ๋ จ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค.
We disable learning rate warmup, set the batch size to 96, the learning rate to 0.001 (with linear decay down to 0), weight decay at 0.1, and finetune for 75,000 steps.
ํ•™์Šต ์†๋„ ์›œ์—…์„ ํ•ด์ œํ•˜๊ณ  ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ 96, ํ•™์Šต ์†๋„๋ฅผ 0.001(์„ ํ˜• ๊ฐ์‡ ๋ฅผ 0์œผ๋กœ ํฌํ•จ), ๋ฌด๊ฒŒ ๊ฐ์‡ ๋ฅผ 0.1๋กœ ์„ค์ •ํ•˜๊ณ  75,000๋‹จ๊ณ„ ๋™์•ˆ ๋ฏธ์„ธ ์กฐ์ •ํ•œ๋‹ค.
Figure 10: The fit result of the scaling experiment with WSD Scheduler. The black dots in a horizontal line denotes the decayed checkpoints in different compute within a same model size.
๊ทธ๋ฆผ 10: WSD Scheduler๋ฅผ ์ด์šฉํ•œ ์Šค์ผ€์ผ๋ง ์‹คํ—˜์˜ ์ ํ•ฉ๋„ ๊ฒฐ๊ณผ. ์ˆ˜ํ‰์„ ์˜ ๊ฒ€์€์ƒ‰ ์ ์€ ๋™์ผํ•œ ๋ชจ๋ธ ํฌ๊ธฐ ๋‚ด์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ ๊ณ„์‚ฐ์—์„œ ๊ฐ์‡ ๋œ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
We set the input length to 32K for fine-tuning and increase to 500K for evaluating. We use a generation temperature of 0.5 and \(top_{p}=0.95\) and set the number of decoding steps to 1024 to generate a summary of each book.
๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ์ž…๋ ฅ ๊ธธ์ด๋ฅผ 32K๋กœ ์„ค์ •ํ•˜๊ณ  ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 500K๋กœ ์ฆ๊ฐ€์‹œํ‚จ๋‹ค. ์šฐ๋ฆฌ๋Š” 0.5์™€ \(top_{p}=0.95\)์˜ ์ƒ์„ฑ ์˜จ๋„๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ๋””์ฝ”๋”ฉ ๋‹จ๊ณ„ ์ˆ˜๋ฅผ 1024๋กœ ์„ค์ •ํ•˜์—ฌ ๊ฐ ์ฑ…์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•œ๋‹ค.
Recent works have shown that Transformer-based (Vaswani et al., 2017) Large Language Models (LLMs) store and retrieve factual information in their parameters to complete simple prompts such as _"The mother of Stevie Wonder is"_(Petroni et al., 2019; Meng et al., 2022; Geva et al., 2021, 2022, 2023; Zhu and Li, 2023). In addition, LLMs have demonstrated remarkable _in-context_ reasoning abilities when the necessary information is explicitly given as part of the input (Wei et al., 2022). For example, models can infer "Lula" as a possible completion of _"The mother of Stevie Wonder is Lula. The singer of 'Superstition' is Stevie Wonder. The mother of the singer of 'Superstition' is"_. These findings raise a question: Do LLMs retrieve factual information stored in their parameters and perform _latent multi-hop reasoning_ when the information to reason from is _not_ given as a part of the input? For instance, when LLMs process the two-hop prompt _"The mother of the singer of 'Superstition' is"_, do they (1) figure out that "the singer of 'Superstition'" refers to Stevie Wonder and (2) use their knowledge of who Stevie Wonder's mother is to complete the prompt?
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด Transformer-based (Vaswani et al., 2017) Large Language Models (LLMs)๋Š” _"The mother of Stevie Wonder is "_(Petroni et al., 2019; Meng et al., 2022; Geva et al., 2021, 2022, 2023; Zhu and Li, 2023)์™€ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋“ค์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•œ๋‹ค. ๋˜ํ•œ, LLM์€ ํ•„์š”ํ•œ ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ์„œ ๋ช…์‹œ์ ์œผ๋กœ ์ฃผ์–ด์งˆ ๋•Œ ํ˜„์ €ํ•œ _in-context_ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ๋‹ค(Wei et al., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋ธ๋“ค์€ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š” ๋ฃฐ๋ผ์ด๋‹ค. '์Šˆํผ์Šคํ‹ฐ์…˜'์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ด๋‹ค. '์Šˆํผ์Šคํ‹ฐ์…˜'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š” _"์˜ ๊ฐ€๋Šฅํ•œ ์™„์„ฑ์œผ๋กœ " ๋ฃฐ๋ผ"๋ฅผ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ๋ฅผ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. LLM์€ ๋งค๊ฐœ ๋ณ€์ˆ˜์— ์ €์žฅ๋œ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  ์ถ”๋ก ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ์ฃผ์–ด์ง€์ง€ ์•Š์€ ๊ฒฝ์šฐ _์ž ์žฌ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก _์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๊นŒ? ์˜ˆ๋ฅผ ๋“ค์–ด LLM์ด 2ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ _"Superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ "_"์ธ ๊ฒฝ์šฐ, ๊ทธ๋“ค์€ (1) "Superstition'์˜ ๊ฐ€์ˆ˜๊ฐ€ ์Šคํ‹ฐ๋น„ ์›๋”๋ฅผ ์ง€์นญํ•œ๋‹ค๋Š” ๊ฒƒ์„ ํŒŒ์•…ํ•˜๊ณ  (2) ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ˆ„๊ตฌ์ธ์ง€์— ๋Œ€ํ•œ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•ฉ๋‹ˆ๊นŒ?
To build a trustworthy benchmark, the grading system needs to be consistent, reproducible, interpretable and well-documented. The WG is committed to being transparent in its entire
์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด ๋“ฑ๊ธ‰ ์‹œ์Šคํ…œ์ด ์ผ๊ด€๋˜๊ณ  ์žฌํ˜„ ๊ฐ€๋Šฅํ•˜๋ฉฐ ํ•ด์„ ๊ฐ€๋Šฅํ•˜๊ณ  ์ž˜ ๋ฌธ์„œํ™”๋˜์–ด์•ผ ํ•œ๋‹ค. WG๋Š” ์ „์ฒด์ ์œผ๋กœ ํˆฌ๋ช…ํ•ด์ง€๋Š” ๋ฐ ์ „๋…ํ•˜๊ณ  ์žˆ๋‹ค.
DoQAWe use the test set of DoQA for the evaluation. Its test set consists of 1200 dialogues with 5394 user-agent turns across Cooking, Travel and Movie domains. Among 5394 user-agent turns, there are 1479 (around 27.4%) unanswerable questions. We use the same strategy as the one mentioned in QuAC dataset to incorporate unanswerable samples into the evaluation.
DoQA ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด DoQA์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” ์š”๋ฆฌ, ์—ฌํ–‰ ๋ฐ ์˜ํ™” ๋„๋ฉ”์ธ์—์„œ 5394๊ฐœ์˜ ์‚ฌ์šฉ์ž ์—์ด์ „ํŠธ๊ฐ€ ํšŒ์ „ํ•˜๋Š” 1200๊ฐœ์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. 5394๊ฐœ์˜ ์‚ฌ์šฉ์ž-์—์ด์ „ํŠธ ํšŒ์ „ ์ค‘ 1479๊ฐœ(์•ฝ 27.4%)์˜ ์‘๋‹ตํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์ด ์žˆ๋‹ค. QuAC ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์–ธ๊ธ‰ํ•œ ๊ฒƒ๊ณผ ๋™์ผํ•œ ์ „๋žต์„ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ตํ•  ์ˆ˜ ์—†๋Š” ์ƒ˜ํ”Œ์„ ํ‰๊ฐ€์— ํ†ตํ•ฉํ•œ๋‹ค.
LLMs have shown impressive capabilities in many downstream tasks in a zero/few-shot setting such as summarization, reading comprehension, translation, and commonsense reasoning Brown et al. (2020); Touvron et al. (2023). To train a LLM, the currently established approach is to employ a tokenizer that splits the training documents into tokens where a token represents a word Bengio et al. (2000), a sub-word Schuster and Nakajima (2012); Sennrich et al. (2015); Wang et al. (2020), or a single character Gao et al. (2020), and each token is represented in the model by an embedding vector that can be further processed.
LLM์€ ์š”์•ฝ, ๋…ํ•ด, ๋ฒˆ์—ญ ๋ฐ ์ƒ์‹ ์ถ”๋ก  Brown ๋“ฑ(2020); Touvron ๋“ฑ(2023)๊ณผ ๊ฐ™์€ ์ œ๋กœ/ํŽ˜์šฐ ์ƒท ํ™˜๊ฒฝ์—์„œ ๋งŽ์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค. LLM์„ ํŠธ๋ ˆ์ด๋‹ํ•˜๊ธฐ ์œ„ํ•ด, ํ˜„์žฌ ํ™•๋ฆฝ๋œ ์ ‘๊ทผ๋ฒ•์€ ํŠธ๋ ˆ์ด๋‹ ๋ฌธ์„œ๋“ค์„ ํ† ํฐ๋“ค๋กœ ๋ถ„ํ• ํ•˜๋Š” ํ† ํฐํ™”๊ธฐ๋ฅผ ์ฑ„์šฉํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ์—ฌ๊ธฐ์„œ ํ† ํฐ์€ ๋‹จ์–ด Bengio ๋“ฑ(2000), ์„œ๋ธŒ์›Œ๋“œ Schuster ๋ฐ Nakajima(2012); Sennrich ๋“ฑ(2015); Wang ๋“ฑ(2020), ๋˜๋Š” ๋‹จ์ผ ๋ฌธ์ž Gao ๋“ฑ(2020)์„ ๋‚˜ํƒ€๋‚ด๊ณ , ๊ฐ๊ฐ์˜ ํ† ํฐ์€ ์ถ”๊ฐ€๋กœ ํ”„๋กœ์„ธ์‹ฑ๋  ์ˆ˜ ์žˆ๋Š” ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ์— ์˜ํ•ด ๋ชจ๋ธ์—์„œ ํ‘œํ˜„๋œ๋‹ค.
(3) **Vicuna** is an open-source chatbot that can generate natural and engaging responses to user queries. It is based on LLaMA and fine-tuned on 70K user-shared conversations collected from ShareGPT, a website where people share their ChatGPT interactions. It is one of the most advanced and versatile open instruction-following models available today. We use the 7B model from FastChat 6.
(3) **Vicuna** ๋Š” ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์— ๋Œ€ ํ•œ ์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋งค๋ ฅ์ ์ธ ์‘๋‹ต์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์˜คํ”ˆ ์†Œ์Šค ์ฑ—๋ด‡์ž…๋‹ˆ๋‹ค. ๊ทธ๊ฒƒ์€ LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ ์‚ฌ๋žŒ๋“ค์ด ChatGPT ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ณต์œ ํ•˜๋Š” ์›น์‚ฌ์ดํŠธ ShareGPT์—์„œ ์ˆ˜์ง‘๋œ 70K ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•œ๋‹ค. ์˜ค๋Š˜๋‚  ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€์žฅ ์ง„๋ณด๋˜๊ณ  ๋‹ค์žฌ๋‹ค๋Šฅํ•œ ๊ฐœ๋ฐฉํ˜• ๋ช…๋ น์–ด ํ›„์† ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” FastChat 6์˜ 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค.
Our work examines the forgetting occurrence during continual pre-training on an existing fine-tuned LLM. Our paper primarily focuses on continual pre-training using the Traditional Chinese corpus. We evaluate the impact of continual pre-training across various dimensions, including output format, knowledge, and reliability. We show that more than straightforward methods are required for resolving this issue. Also, we observe an increased prominence of the repetition problem in models that tend to generate Traditional Chinese outputs. Lastly, despite continual pre-training, our findings suggest that the model's knowledge remains unaffected while its reliability declines.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๊ธฐ์กด์˜ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘ ๋ง๊ฐ ๋ฐœ์ƒ์„ ์กฐ์‚ฌํ•œ๋‹ค. ์šฐ๋ฆฌ์˜ ๋…ผ๋ฌธ์€ ์ฃผ๋กœ ์ค‘๊ตญ ์ „ํ†ต ๋ง๋ญ‰์น˜๋ฅผ ์ด์šฉํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ถœ๋ ฅ ํ˜•์‹, ์ง€์‹ ๋ฐ ์‹ ๋ขฐ์„ฑ์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ฐจ์›์— ๊ฑธ์ณ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹จ์ˆœํ•œ ๋ฐฉ๋ฒ• ์ด์ƒ์˜ ๋ฐฉ๋ฒ•์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค. ๋˜ํ•œ ์ „ํ†ต์ ์ธ ์ค‘๊ตญ์–ด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๋ชจ๋ธ์—์„œ ๋ฐ˜๋ณต ๋ฌธ์ œ์˜ ์ค‘์š”์„ฑ์ด ์ฆ๊ฐ€ํ–ˆ์Œ์„ ๊ด€์ฐฐํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹์ด ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š๋Š” ๋ฐ˜๋ฉด ์‹ ๋ขฐ์„ฑ์€ ๊ฐ์†Œํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•œ๋‹ค.
**10% Steps are Enough.** From the two stage training perspective, shortening the decay stage will greatly benefit the fast test of different model checkpoints of stable training. Therefore, we conduct experiments that start from the same stable training checkpoints and have different decay steps. Also shown in Figure 5, among all three stable training checkpoints in 40N, 60N, and 80N training data, having a decay of 10% of the total tokens is sufficient to achieve the best results, while a decay of 2.5% of total tokens falls short. Therefore, in the subsequent training experiments, we use a decay of about 10% to ensure full convergence.
**10% ๋‹จ๊ณ„๋Š” ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.* * ๋‘ ๋‹จ๊ณ„ ํ›ˆ๋ จ ๊ด€์ ์—์„œ ๋ถ•๊ดด ๋‹จ๊ณ„๋ฅผ ๋‹จ์ถ•ํ•˜๋ฉด ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์˜ ๋น ๋ฅธ ํ…Œ์ŠคํŠธ์— ํฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ๋™์ผํ•œ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ์ฒดํฌํฌ์ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ๊ฐ์‡  ๋‹จ๊ณ„๊ฐ€ ๋‹ค๋ฅธ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•œ๋‹ค. ๊ทธ๋ฆผ 5์—์„œ๋„ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด 40N, 60N, 80N ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ์„ธ ๊ฐ€์ง€ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ์ฒดํฌํฌ์ธํŠธ ์ค‘ ์ด ํ† ํฐ์˜ 10%์˜ ๊ฐ์‡ ๋Š” ์ตœ์ƒ์˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป๊ธฐ์— ์ถฉ๋ถ„ํ•˜์ง€๋งŒ ์ด ํ† ํฐ์˜ 2.5%์˜ ๊ฐ์‡ ๋Š” ๋ถ€์กฑํ•œ๋‹ค. ๋”ฐ๋ผ์„œ, ํ›„์† ํ›ˆ๋ จ ์‹คํ—˜์—์„œ, ์šฐ๋ฆฌ๋Š” ์™„์ „ํ•œ ์ˆ˜๋ ด์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์•ฝ 10%์˜ ๊ฐ์‡ ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.
\(\bullet\)_Reward model based approaches._ The reward model in RLHF has been trained to measure the alignment degree on the responses of LLMs. It is straightforward to leverage existing reward models to select high-quality responses as alignment data for subsequent fine-tuning. Based on this idea, RAFT [373] adopts reward models trained on human preference data to rank the responses of LLMs and collect those with higher rewards for supervised fine-tuning. In addition, the reward model can be also used to score model responses and assign them to different quality groups. Quark [384] sorts the responses of LLMs into different quantiles based on the reward scores. Each quantile is attached with a special reward token to represent the reward level of the quantile. Conditioned on the highest-reward tokens, LLMs are subsequently prompted to generate high-quality responses. Given an initial answer and the corresponding human feedback, ILF [385] first adopts LLMs to generate refined answers, then utilizes the reward model to select the answer that best matches the feedback for further training. As valuable resources for aligning LLMs, several reward models have been released, including DeBERTabase/large/xlarge from OpenAssistant27, Moss-7B from Fudan28, and Flan-T5-xl from Stanford29.
\(\bullet\)_๋ณด์ƒ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•._ RLHF์˜ ๋ณด์ƒ ๋ชจ๋ธ์€ LLM์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์ •๋ ฌ ์ •๋„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ๋‹ค. ๊ธฐ์กด ๋ณด์ƒ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ๋ฐ˜์‘์„ ํ›„์† ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•œ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ๊ฐ„๋‹จํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•„์ด๋””์–ด์— ๊ธฐ์ดˆํ•˜์—ฌ, RAFT[373]๋Š” ์ธ๊ฐ„ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํŠธ๋ ˆ์ด๋‹๋œ ๋ณด์ƒ ๋ชจ๋ธ๋“ค์„ ์ฑ„ํƒํ•˜์—ฌ LLM๋“ค์˜ ์‘๋‹ต๋“ค์˜ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๊ณ  ๊ฐ๋…๋œ ๋ฏธ์„ธ ์กฐ์ •์— ๋Œ€ํ•ด ๋” ๋†’์€ ๋ณด์ƒ์„ ๊ฐ–๋Š” ๊ฒƒ๋“ค์„ ์ˆ˜์ง‘ํ•œ๋‹ค. ๋˜ํ•œ, ๋ณด์ƒ ๋ชจ๋ธ์€ ๋ชจ๋ธ ์‘๋‹ต๋“ค์„ ์Šค์ฝ”์–ด๋งํ•˜๊ณ  ์ƒ์ดํ•œ ํ’ˆ์งˆ ๊ทธ๋ฃน๋“ค์— ํ• ๋‹นํ•˜๊ธฐ ์œ„ํ•ด ๋˜ํ•œ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค. ์ฟผํฌ[384]๋Š” ๋ณด์ƒ ์ ์ˆ˜์— ๊ธฐ์ดˆํ•˜์—ฌ LLM์˜ ์‘๋‹ต์„ ์ƒ์ดํ•œ ๋ถ„์œ„์ˆ˜๋กœ ์ •๋ ฌํ•œ๋‹ค. ๊ฐ ๋ถ„์œ„์—๋Š” ๋ถ„์œ„์˜ ๋ณด์ƒ ์ˆ˜์ค€์„ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ๋ณด์ƒ ํ† ํฐ์ด ๋ถ€์ฐฉ๋˜์–ด ์žˆ๋‹ค. ์ตœ๊ณ  ๋ฆฌ์›Œ๋“œ ํ† ํฐ์— ๋Œ€ํ•ด ์กฐ๊ฑดํ™”๋œ LLM์€ ํ›„์†์ ์œผ๋กœ ๊ณ ํ’ˆ์งˆ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋œ๋‹ค. ์ดˆ๊ธฐ ๋‹ต๋ณ€ ๋ฐ ๋Œ€์‘ํ•˜๋Š” ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์ด ์ฃผ์–ด์ง€๋ฉด, ILF[385]๋Š” ๋จผ์ € LLMs๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ์ •์ œ๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ๋ณด์ƒ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ์ถ”๊ฐ€ ํŠธ๋ ˆ์ด๋‹์„ ์œ„ํ•ด ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐ€์žฅ ์ผ์น˜ํ•˜๋Š” ๋‹ต๋ณ€์„ ์„ ํƒํ•œ๋‹ค. LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•œ ๊ท€์ค‘ํ•œ ์ž์›์œผ๋กœ OpenAssistant27์˜ DeBERTabase/large/xlarge, Fudan28์˜ Moss-7B, Stanford29์˜ Flan-T5-xl ๋“ฑ ์—ฌ๋Ÿฌ ๋ณด์ƒ ๋ชจ๋ธ์ด ์ถœ์‹œ๋˜์—ˆ๋‹ค.
An example of how these ethical theories would apply to a specific ethical dilemma is the ethical use of artificial intelligence. Virtue ethics would suggest that the best way to ensure ethical AI use is to prioritize the development of moral virtues such as transparency, accountability, and fairness. Deontological ethics, on the other hand, would dictate that AI systems must be designed and used in a way that adheres to a set of pre-defined moral rules, regardless of the consequences.
์ด๋Ÿฌํ•œ ์œค๋ฆฌ ์ด๋ก ์ด ํŠน์ • ์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ์— ์–ด๋–ป๊ฒŒ ์ ์šฉ๋  ๊ฒƒ์ธ๊ฐ€์— ๋Œ€ํ•œ ์˜ˆ๋Š” ์ธ๊ณต ์ง€๋Šฅ์˜ ์œค๋ฆฌ์  ํ™œ์šฉ์ด๋‹ค. ๋• ์œค๋ฆฌ๋Š” ์œค๋ฆฌ์  AI ์‚ฌ์šฉ์„ ๋ณด์žฅํ•˜๋Š” ๊ฐ€์žฅ ์ข‹์€ ๋ฐฉ๋ฒ•์€ ํˆฌ๋ช…์„ฑ, ์ฑ…์ž„์„ฑ, ๊ณต์ •์„ฑ๊ณผ ๊ฐ™์€ ๋„๋•์  ๋•๋ชฉ์˜ ๊ฐœ๋ฐœ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด๋ผ๊ณ  ์ œ์•ˆํ•  ๊ฒƒ์ด๋‹ค. ๋ฐ˜๋ฉด์— ์˜๋ฌด๋ก ์  ์œค๋ฆฌ๋Š” AI ์‹œ์Šคํ…œ์ด ๊ฒฐ๊ณผ์— ๊ด€๊ณ„์—†์ด ๋ฏธ๋ฆฌ ์ •์˜๋œ ์ผ๋ จ์˜ ๋„๋•์  ๊ทœ์น™์„ ์ค€์ˆ˜ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์„ค๊ณ„๋˜๊ณ  ์‚ฌ์šฉ๋˜์–ด์•ผ ํ•œ๋‹ค๊ณ  ์ง€์‹œํ•  ๊ฒƒ์ด๋‹ค.
Multiple epochs.Instead of looking for "unique" tokens to make up a trillion-scale pretraining dataset, one could simply repeat data over multiple epochs. Popular models like OPT and NeoX-20B do this for up to 2 epochs, and most curated datasets upsample corpora 2-5 times. However, Hernandez et al. (2022) has recently shown that models with 100B+ parameters may be sensitive to even just a few epochs. Orthogonal to our work lies a line of research exploring tradeoffs in the data-constrained regime: can deduplication help sustain more epochs? Are multiple epochs on higher quality data better than a one epoch on lower quality data? See Appendix E.3 for a more in-depth discussion.
์—ฌ๋Ÿฌ ์‹œ๋Œ€.์กฐ ๊ทœ๋ชจ์˜ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด "๊ณ ์œ ํ•œ" ํ† ํฐ์„ ์ฐพ๋Š” ๋Œ€์‹  ์—ฌ๋Ÿฌ ์‹œ๋Œ€์— ๊ฑธ์ณ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. OPT ๋ฐ NeoX-20B์™€ ๊ฐ™์€ ์ธ๊ธฐ ์žˆ๋Š” ๋ชจ๋ธ์€ ์ตœ๋Œ€ 2๊ฐœ์˜ ์—ํญ์— ๋Œ€ํ•ด ์ด ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ ๋Œ€๋ถ€๋ถ„์˜ ํ๋ ˆ์ดํŠธ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 2-5ํšŒ ๋ง๋ญ‰์น˜๋ฅผ ์—…์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Hernandez et al.(2022)์€ ์ตœ๊ทผ 100B+ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์ด ๋ช‡ ๋ฒˆ์˜ ์—ํฌํฌ์—๋„ ๋ฏผ๊ฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ž‘์—…๊ณผ ์ง๊ตํ•˜๋Š” ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ์ œ์•ฝ ์ฒด์ œ์—์„œ ์ ˆ์ถฉ์ ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฐ๊ตฌ ๋ผ์ธ์— ์žˆ๋‹ค: ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ๋” ๋งŽ์€ ์‹œ๋Œ€๋ฅผ ์ง€์†ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š”๊ฐ€? ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์˜ ์—ฌ๋Ÿฌ ์—ํฌํฌ๊ฐ€ ์ €ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์˜ ํ•œ ์—ํฌํฌ๋ณด๋‹ค ๋‚˜์€๊ฐ€์š”? ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก E.3์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
On Friday, the mechanic repaired 12 car tires and didn't repair any truck tires. Therefore, the revenue earned on Friday can be calculated as follows:
๊ธˆ์š”์ผ์— ์ •๋น„์‚ฌ๋Š” ์ž๋™์ฐจ ํƒ€์ด์–ด 12๊ฐœ๋ฅผ ์ˆ˜๋ฆฌํ–ˆ๊ณ  ํŠธ๋Ÿญ ํƒ€์ด์–ด๋Š” ์ˆ˜๋ฆฌํ•˜์ง€ ์•Š์•˜๋‹ค. ๋”ฐ๋ผ์„œ ๊ธˆ์š”์ผ์— ๋ฒŒ์–ด๋“ค์ธ ์ˆ˜์ต์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐ๋  ์ˆ˜ ์žˆ๋‹ค:
Quantization is a popular strategy for decreasing a model's memory footprint and speeding up inference. In addition to these benefits, using lower-precision number representations also effectively adds noise. As noted in Section 8.2, adding noise to the output logits could prevent our attack. A natural question that follows is, does quantization add sufficient noise to make our attack ineffective or more difficult to carry out?
์–‘์žํ™”๋Š” ๋ชจ๋ธ์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ณต๊ฐ„์„ ์ค„์ด๊ณ  ์ถ”๋ก ์„ ๋น ๋ฅด๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•œ ์ธ๊ธฐ ์žˆ๋Š” ์ „๋žต์ด๋‹ค. ์ด๋Ÿฌํ•œ ์ด์  ์™ธ์—๋„ ๋” ๋‚ฎ์€ ์ •๋ฐ€๋„ ์ˆ˜ ํ‘œํ˜„์„ ์‚ฌ์šฉํ•˜๋ฉด ๋…ธ์ด์ฆˆ๋„ ํšจ๊ณผ์ ์œผ๋กœ ์ถ”๊ฐ€๋œ๋‹ค. ์„น์…˜ 8.2์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ์ถœ๋ ฅ ๋กœ์ง“์— ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ๊ณต๊ฒฉ์„ ๋ฐฉ์ง€ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋‹ค์Œ์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ์งˆ๋ฌธ์€ ์–‘์žํ™”๊ฐ€ ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์„ ๋น„ํšจ์œจ์ ์œผ๋กœ ๋งŒ๋“ค๊ฑฐ๋‚˜ ์ˆ˜ํ–‰ํ•˜๊ธฐ ๋” ์–ด๋ ต๊ฒŒ ๋งŒ๋“ค๊ธฐ์— ์ถฉ๋ถ„ํ•œ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š๋ƒ์ด๋‹ค.
**Action Space**: Action space refers to the set of possible actions that can be performed by the agent. In general, we can roughly divide these actions into two classes: (1) external tools and (2) internal knowledge of the LLMs. In the following, we introduce these actions more in detail.
**์ž‘์—… ๊ณต๊ฐ„**: ์ž‘์—… ๊ณต๊ฐ„์€ ์—์ด์ „ํŠธ์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€๋Šฅํ•œ ์ž‘์—… ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ํ–‰๋™์„ (1) ์™ธ๋ถ€ ๋„๊ตฌ์™€ (2) LLM์— ๋Œ€ํ•œ ๋‚ด๋ถ€ ์ง€์‹์˜ ๋‘ ๊ฐ€์ง€ ํด๋ž˜์Šค๋กœ ๋Œ€๋ณ„ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ดํ•˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํ–‰์œ„๋ฅผ ์ข€ ๋” ๊ตฌ์ฒด์ ์œผ๋กœ ์†Œ๊ฐœํ•˜๊ณ ์ž ํ•œ๋‹ค.
LanguagesWe adopt \(6\) languages including English (En), German (De), French (Fr), Chinese (Zh), Spanish (Es), and Russian (Ru). These languages were selected because their training corpora represent over \(0.1\%\) of the total pre-training data for Llama2. Furthermore, both Vicuna and Mistral are designed to tackle problems across these languages effectively, and their performance in these languages is much higher than in other low-resource languages such as Thai, Swahili, and Greek.
์šฐ๋ฆฌ๋Š” ์˜์–ด(En), ๋…์ผ์–ด(De), ํ”„๋ž‘์Šค์–ด(Fr), ์ค‘๊ตญ์–ด(Zh), ์ŠคํŽ˜์ธ์–ด(Es), ๋Ÿฌ์‹œ์•„์–ด(Ru)๋ฅผ ํฌํ•จํ•œ \(6\) ์–ธ์–ด๋ฅผ ์ฑ„ํƒํ•œ๋‹ค. ์ด๋“ค ์–ธ์–ด๋Š” Llama2์— ๋Œ€ํ•œ ์ „์ฒด ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ \(0.1\%\) ์ด์ƒ์„ ๋‚˜ํƒ€๋‚ด๊ธฐ ๋•Œ๋ฌธ์— ์„ ํƒ๋˜์—ˆ๋‹ค. ๋˜ํ•œ, Vicuna์™€ Mistral์€ ์ด๋“ค ์–ธ์–ด๋“ค ์‚ฌ์ด์˜ ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, ์ด๋“ค ์–ธ์–ด์—์„œ์˜ ์„ฑ๋Šฅ์€ ํƒœ๊ตญ์–ด, ์Šค์™€ํž๋ฆฌ์–ด, ๊ทธ๋ฆฌ์Šค์–ด์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ์ €์ž์› ์–ธ์–ด๋“ค์— ๋น„ํ•ด ํ›จ์”ฌ ๋†’๋‹ค.
3. Model Architecture: Choose a suitable model architecture that can handle a wide range of tasks and has the ability to learn from different types of data. InstructGPT is a good choice for this task since it is designed for natural language processing.
3. ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜: ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๊ณ  ์ƒ์ดํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–๋Š” ์ ํ•ฉํ•œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ ํƒํ•œ๋‹ค. InstructGPT๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ด ์ž‘์—…์— ์ข‹์€ ์„ ํƒ์ด๋‹ค.
We pulled Olympics records pages from Wikipedia.org across 9 sports: Athletics, weightlifting, swimming, archery, track cycling, rowing, shooting, short track speed skating, and speed skating. Records are extracted in a table format, from which questions are generated for each record entry. In total, after filtering, we extracted 192 unique questions and answers.
์šฐ๋ฆฌ๋Š” ์œ„ํ‚คํ”ผ๋””์•„.org์—์„œ ์œก์ƒ, ์—ญ๋„, ์ˆ˜์˜, ์–‘๊ถ, ํŠธ๋ž™ ์‚ฌ์ดํด, ์กฐ์ •, ์‚ฌ๊ฒฉ, ์‡ผํŠธํŠธ๋ž™ ์Šคํ”ผ๋“œ ์Šค์ผ€์ดํŒ…, ์Šคํ”ผ๋“œ ์Šค์ผ€์ดํŒ… ๋“ฑ 9๊ฐœ ์ข…๋ชฉ์— ๊ฑธ์ณ ์˜ฌ๋ฆผํ”ฝ ๊ธฐ๋ก ํŽ˜์ด์ง€๋ฅผ ๋ฝ‘์•˜๋‹ค. ๋ ˆ์ฝ”๋“œ๋Š” ํ…Œ์ด๋ธ” ํ˜•์‹์œผ๋กœ ์ถ”์ถœ๋˜๋ฉฐ, ์ด๋กœ๋ถ€ํ„ฐ ๊ฐ ๋ ˆ์ฝ”๋“œ ํ•ญ๋ชฉ์— ๋Œ€ํ•œ ์งˆ๋ฌธ์ด ์ƒ์„ฑ๋œ๋‹ค. ์ „์ฒด์ ์œผ๋กœ ํ•„ํ„ฐ๋ง ํ›„ 192๊ฐœ์˜ ๊ณ ์œ ํ•œ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ถ”์ถœํ–ˆ๋‹ค.
Note that this algorithm still requires that the dataset itself fits in memory (so that we can efficiently index in arbitrary positions), but we do not need to fit the entire suffix array into memory. This is fortunate since our suffix array requires an \(8\times\) space overhead. For example, the suffix array for the \(350\)GB C4 is \(1.5\)TB.
์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ฐ์ดํ„ฐ์„ธํŠธ ์ž์ฒด๊ฐ€ (์ž„์˜์˜ ์œ„์น˜์—์„œ ํšจ์œจ์ ์œผ๋กœ ์ธ๋ฑ์‹ฑํ•  ์ˆ˜ ์žˆ๋„๋ก) ๋ฉ”๋ชจ๋ฆฌ์— ์ ํ•ฉํ•ด์•ผ ํ•˜์ง€๋งŒ, ์ „์ฒด ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์„ ๋ฉ”๋ชจ๋ฆฌ์— ์ ํ•ฉ์‹œํ‚ฌ ํ•„์š”๋Š” ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•œ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์€ \(8\times\) ๊ณต๊ฐ„ ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๊ฒƒ์€ ๋‹คํ–‰์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(350\)GB C4์— ๋Œ€ํ•œ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์€ \(1.5\)TB์ž…๋‹ˆ๋‹ค.
Recently, the AI landscape has borne witness to the emergence of numerous open-source LLMs, characterized by publicly accessible model codes and weight parameters. EleutherAI has contributed GPT-NeoX-20B [54] and GPT-J-6B [48]. BigScience has introduced BLOOM [55]. Similarly, Meta has made strides by releasing OPT [53], Llama 1 [4], Llama 2 [20], and GAL [14]. Tsinghua University has unveiled GLM-130B and ChatGLM [45]. TII has facilitated the release of Falcon [21]. Additionally, LLMs such as Baichuan9 and Qwen10 have also surfaced. Presently, Llama assumes a pivotal role as the foundational model for supervised fine-tuning, ushering in the emergence of several extremely remarkable models, including Alpaca [22], Vicuna [23], Guanaco [56], WizardLM [24], and Orca [57], RFT [16] etc.
์ตœ๊ทผ AI ๊ฒฝ๊ด€์€ ๊ณต๊ฐœ์ ์œผ๋กœ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ ์ฝ”๋“œ์™€ ๊ฐ€์ค‘์น˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ ํŠน์ง•์ง€์–ด์ง€๋Š” ์ˆ˜๋งŽ์€ ์˜คํ”ˆ ์†Œ์Šค LLM์ด ์ถœํ˜„ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉ๊ฒฉํ–ˆ๋‹ค. EleutherAI๋Š” GPT-NeoX-20B[54]์™€ GPT-J-6B[48]์— ๊ธฐ์—ฌํ–ˆ๋‹ค. ๋น…์‚ฌ์ด์–ธ์Šค๋Š” BLOOM[55]์„ ๋„์ž…ํ–ˆ๋‹ค. ์œ ์‚ฌํ•˜๊ฒŒ, ๋ฉ”ํƒ€๋Š” OPT[53], Llama 1[4], Llama 2[20], ๋ฐ GAL[14]์„ ๋ฐฉ์ถœํ•จ์œผ๋กœ์จ ์ง„๋ณด๋ฅผ ์ด๋ฃจ์—ˆ๋‹ค. ์นญํ™”๋Œ€๋Š” GLM-130B์™€ ChatGLM[45]๋ฅผ ๊ณต๊ฐœํ–ˆ๋‹ค. TII๋Š” ํŒ”์ฝ˜์˜ ๋ฐฉ์ถœ์„ ์ด‰์ง„ํ–ˆ๋‹ค[21]. ๋˜ ๋ฐ”์ด์ดจ9, ํ์›ฌ10 ๋“ฑ LLM๋„ ์ˆ˜๋ฉด ์œ„๋กœ ๋– ์˜ฌ๋ž๋‹ค. ํ˜„์žฌ Llama๋Š” ๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•œ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์ค‘์ถ”์ ์ธ ์—ญํ• ์„ ํ•˜๋ฉฐ Alpaca[22], Vicuna[23], Guanaco[56], WizardLM[24], Orca[57], RFT[16] ๋“ฑ์„ ํฌํ•จํ•œ ๋ช‡ ๊ฐ€์ง€ ๋งค์šฐ ๋†€๋ผ์šด ๋ชจ๋ธ์˜ ์ถœํ˜„์„ ์˜ˆ๊ณ ํ•œ๋‹ค.
**LLM-enhanced Recommendation Models.** In addition to instructing LLMs to directly provide recommendations, researchers also propose leveraging the universal knowledge encoded in LLMs to improve traditional recommender systems. Existing approaches in this line can be divided into three main categories. The first category employs LLMs to infer users' potential intention from their historical interaction data. Furthermore, traditional recommendation/search models employ the inferred intentions to improve the retrieval of relevant items [812, 813]. Additionally, several studies explore the use of LLMs as feature encoders. They employ LLMs to encode the side information of items and users (_e.g._, item's descriptions and user's reviews), thus deriving more informative representations of users and items. These representations are then fed into traditional recommender systems as augmented input [814, 815]. As another alternative approach, several studies [816, 817] adopt a distillation-like way to transfer LLM's capacities (_e.g._, semantic encoding) to improve traditional recommenders (_i.e._, small models). Specially, they align the hidden states of LLMs and traditional recommendation models via joint training. After training, since only the enhanced small model will be deployed online, it can avoid the huge overhead of LLMs in online service.
**LLM ๊ฐ•ํ™” ์ถ”์ฒœ ๋ชจ๋ธ** LLM์ด ๊ถŒ์žฅ ์‚ฌํ•ญ์„ ์ง์ ‘ ์ œ๊ณตํ•˜๋„๋ก ์ง€์‹œํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ์—ฐ๊ตฌ์ž๋Š” LLM์— ์ธ์ฝ”๋”ฉ๋œ ๋ณดํŽธ์ ์ธ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ์ „ํ†ต์ ์ธ ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ๋„ ์ œ์•ˆํ•œ๋‹ค. ์ด ๋ผ์ธ์˜ ๊ธฐ์กด ์ ‘๊ทผ ๋ฐฉ์‹์€ ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฒ”์ฃผ๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ๊ณผ๊ฑฐ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ์—์„œ ์‚ฌ์šฉ์ž์˜ ์ž ์žฌ์  ์˜๋„๋ฅผ ์ถ”๋ก ํ•œ๋‹ค. ๋˜ํ•œ, ์ „ํ†ต์ ์ธ ์ถ”์ฒœ/๊ฒ€์ƒ‰ ๋ชจ๋ธ์€ ๊ด€๋ จ ํ•ญ๋ชฉ์˜ ๊ฒ€์ƒ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๋ก ๋œ ์˜๋„๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค[812, 813]. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ LLM์„ ํŠน์ง• ์ธ์ฝ”๋”๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•œ๋‹ค. ์ด๋“ค์€ LLMs์„ ์‚ฌ์šฉํ•˜์—ฌ ์•„์ดํ…œ ๋ฐ ์‚ฌ์šฉ์ž์˜ ๋ถ€๊ฐ€ ์ •๋ณด(_e.g._, ์•„์ดํ…œ์˜ ์„ค๋ช… ๋ฐ ์‚ฌ์šฉ์ž์˜ ๋ฆฌ๋ทฐ)๋ฅผ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ์‚ฌ์šฉ์ž ๋ฐ ์•„์ดํ…œ์˜ ๋ณด๋‹ค ์œ ์ตํ•œ ํ‘œํ˜„์„ ์œ ๋„ํ•œ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ด๋Ÿฌํ•œ ํ‘œํ˜„๋“ค์€ ์ฆ๊ฐ• ์ž…๋ ฅ์œผ๋กœ์„œ ์ „ํ†ต์ ์ธ ์ถ”์ฒœ ์‹œ์Šคํ…œ๋“ค๋กœ ๊ณต๊ธ‰๋œ๋‹ค[814, 815]. ๋‹ค๋ฅธ ๋Œ€์•ˆ์  ์ ‘๊ทผ๋ฒ•์œผ๋กœ์„œ, ๋ช‡๋ช‡ ์—ฐ๊ตฌ๋“ค[816, 817]์€ ์ „ํ†ต์ ์ธ ์ถ”์ฒœ์ž๋“ค(_i.e._, ์†Œํ˜• ๋ชจ๋ธ๋“ค)์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM๋“ค์˜ ์šฉ๋Ÿ‰๋“ค(_e.g._, ์‹œ๋งจํ‹ฑ ์ธ์ฝ”๋”ฉ)์„ ์ „๋‹ฌํ•˜๋Š” ์ฆ๋ฅ˜-์œ ์‚ฌ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•œ๋‹ค. ํŠนํžˆ, LLM๊ณผ ์ „ํ†ต์ ์ธ ์ถ”์ฒœ ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ํ•ฉ๋™ ํ›ˆ๋ จ์„ ํ†ตํ•ด ์ •๋ ฌํ•œ๋‹ค. ๊ต์œก ํ›„์—๋Š” ํ–ฅ์ƒ๋œ ์†Œํ˜• ๋ชจ๋ธ๋งŒ ์˜จ๋ผ์ธ์œผ๋กœ ๋ฐฐํฌ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์˜จ๋ผ์ธ ์„œ๋น„์Šค์—์„œ LLM์˜ ๋ง‰๋Œ€ํ•œ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํ”ผํ•  ์ˆ˜ ์žˆ๋‹ค.
State Space Model has been widely studied and applied in many applications, however, the research in this direction is still in its early stages. To help the readers quickly grasp the frontiers, this paper puts forward several research points worthy of attention.
State Space Model์€ ๋งŽ์€ ์‘์šฉ ๋ถ„์•ผ์—์„œ ๋„๋ฆฌ ์—ฐ๊ตฌ๋˜๊ณ  ์ ์šฉ๋˜์—ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ๋ฐฉํ–ฅ์œผ๋กœ์˜ ์—ฐ๊ตฌ๋Š” ์•„์ง ์ดˆ๊ธฐ ๋‹จ๊ณ„์— ์žˆ๋‹ค. ๋…์ž๋“ค์ด ํ”„๋ก ํ‹ฐ์–ด๋ฅผ ์‹ ์†ํ•˜๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋ช‡ ๊ฐ€์ง€ ์—ฐ๊ตฌ ํฌ์ธํŠธ๋ฅผ ์ œ์‹œํ•œ๋‹ค.
provides a mathematical representation of how the performance (\(P\)) of machine learning models scales with respect to three critical factors: model size (\(N\)), data size (\(D\)), and computational budget (\(C\)). Each of these components plays a crucial role in the model's ability to learn from data and perform tasks accurately. Here is a detailed explanation of each parameter within the formula:
๊ธฐ๊ณ„ ํ•™์Šต ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ(\(P\))์ด ๋ชจ๋ธ ํฌ๊ธฐ(\(N\)), ๋ฐ์ดํ„ฐ ํฌ๊ธฐ(\(D\)), ๊ณ„์‚ฐ ์˜ˆ์‚ฐ(\(C\))์˜ ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์š”์†Œ์— ๋Œ€ํ•ด ์–ด๋–ป๊ฒŒ ์Šค์ผ€์ผ๋ง๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์ˆ˜ํ•™์  ํ‘œํ˜„์„ ์ œ๊ณตํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šตํ•˜๊ณ  ์ž‘์—…์„ ์ •ํ™•ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค. ๋‹ค์Œ์€ ์ˆ˜์‹ ๋‚ด์˜ ๊ฐ ํŒŒ๋ผ๋ฏธํ„ฐ์— ๋Œ€ํ•œ ์ƒ์„ธํ•œ ์„ค๋ช…์ด๋‹ค:
Specialized Knowledge TasksThe key message from the results in Table 2 is that synthetic data can not impart 'new knowledge'. It can only help pre-train faster, which was also the premise of our work. In particular, we note several key findings:
์ „๋ฌธํ™”๋œ ์ง€์‹ ๊ณผ์ œ ํ‘œ 2์˜ ๊ฒฐ๊ณผ์—์„œ ๋‚˜์˜จ ํ•ต์‹ฌ ๋ฉ”์‹œ์ง€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ '์ƒˆ๋กœ์šด ์ง€์‹'์„ ์ „๋‹ฌํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ๊ทธ๊ฒƒ์€ ์šฐ๋ฆฌ์˜ ์ž‘์—…์˜ ์ „์ œ์ด๊ธฐ๋„ ํ–ˆ๋˜ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ๋” ๋น ๋ฅด๊ฒŒ ๋„์šธ ์ˆ˜ ์žˆ์„ ๋ฟ์ด๋‹ค. ํŠนํžˆ, ์šฐ๋ฆฌ๋Š” ๋ช‡ ๊ฐ€์ง€ ์ฃผ์š” ๋ฐœ๊ฒฌ์— ์ฃผ๋ชฉํ•œ๋‹ค.
Scaling up the generation of synthetic math data is a straightforward process, but ensuring the correctness of the generated math remains a significant challenge for practitioners. AlphaGeometry (Trinh et al., 2024) is a recent attempt to address this issue by training a neural model using 100 million synthetic data points. The model proposes solutions and guides a symbolic deduction engine in verifying the correctness of each branch when solving complex geometry problems. By combining the power of synthetic data with a rigorous verification process, AlphaGeometry achieves a problem-solving ability comparable to that of a human Olympiad gold medalist, demonstrating the potential of this approach in tackling complex mathematical reasoning tasks.
ํ•ฉ์„ฑ ์ˆ˜ํ•™ ๋ฐ์ดํ„ฐ์˜ ์ƒ์„ฑ์„ ํ™•๋Œ€ํ•˜๋Š” ๊ฒƒ์€ ๊ฐ„๋‹จํ•œ ๊ณผ์ •์ด์ง€๋งŒ, ์ƒ์„ฑ๋œ ์ˆ˜ํ•™์˜ ์ •ํ™•์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์‹ค๋ฌด์ž๋“ค์—๊ฒŒ ์—ฌ์ „ํžˆ ์ค‘์š”ํ•œ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ๋‹ค. AlphaGeometry (Trinh et al., 2024)๋Š” 1์–ต ๊ฐœ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์„ ํŠธ๋ ˆ์ด๋‹ํ•จ์œผ๋กœ์จ ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋ ค๋Š” ์ตœ๊ทผ์˜ ์‹œ๋„์ด๋‹ค. ์ด ๋ชจ๋ธ์€ ๋ณต์žกํ•œ ๊ธฐํ•˜ํ•™ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๋•Œ ๊ฐ ๋ถ„๊ธฐ์˜ ์ •ํ™•์„ฑ์„ ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ์žˆ์–ด ์†”๋ฃจ์…˜์„ ์ œ์•ˆํ•˜๊ณ  ๊ธฐํ˜ธ ์ถ”๋ก  ์—”์ง„์„ ์•ˆ๋‚ดํ•œ๋‹ค. ์•ŒํŒŒ๊ธฐํ•˜ํ•™์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํž˜์„ ์—„๊ฒฉํ•œ ๊ฒ€์ฆ ๊ณผ์ •๊ณผ ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ ์ธ๊ฐ„ ์˜ฌ๋ฆผํ”ผ์•„๋“œ ๊ธˆ๋ฉ”๋‹ฌ๋ฆฌ์ŠคํŠธ์— ๋ฒ„๊ธˆ๊ฐ€๋Š” ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ๋ณต์žกํ•œ ์ˆ˜ํ•™์  ์ถ”๋ก  ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ด ์ ‘๊ทผ๋ฒ•์˜ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ค€๋‹ค.
Jurning discover a hidden on his brain. Prison house filled with action-team from school and the group size.
๋ฒ„๋‹์€ ๊ทธ์˜ ๋‡Œ์— ์ˆจ๊ฒจ์ง„ ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•œ๋‹ค. ๊ต๋„์†Œ์—๋Š” ํ•™๊ต์—์„œ ์˜จ ์•ก์…˜ ํŒ€๊ณผ ๋‹จ์ฒด ๊ทœ๋ชจ๊ฐ€ ๊ฐ€๋“ํ•˜๋‹ค.
Figure 3: **Left.** Win rates computed by GPT-4 for Anthropic-HH one-step dialogue; DPO is the only method that improves over chosen summaries in the Anthropic-HH test set. **Right.** Win rates for different sampling temperatures over the course of training. DPOโ€™s improvement over the dataset labels is fairly stable over the course of training for different sampling temperatures.
๊ทธ๋ฆผ 3: **์™ผ์ชฝ.** Anthropic-HH 1๋‹จ๊ณ„ ๋Œ€ํ™”์— ๋Œ€ํ•ด GPT-4์—์„œ ๊ณ„์‚ฐ๋œ ์Šน๋ฅ ; DPO๋Š” Anthropic-HH ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ์„ ํƒํ•œ ์š”์•ฝ๋ณด๋‹ค ๊ฐœ์„ ๋˜๋Š” ์œ ์ผํ•œ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. **๋งž์•„์š”.** ๊ต์œก ๊ณผ์ •์—์„œ ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ๋ง ์˜จ๋„์— ๋Œ€ํ•œ ์Šน๋ฆฌ์œจ์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ผ๋ฒจ์— ๋Œ€ํ•œ DPO์˜ ๊ฐœ์„ ์€ ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ๋ง ์˜จ๋„์— ๋Œ€ํ•œ ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ์ƒ๋‹นํžˆ ์•ˆ์ •์ ์ด๋‹ค.
Instruction-following LMs.A series of works have found evidence that vanilla LMs can be effective at following general language instructions if tuned with annotated "instructional" data--datasets containing language instructional commands and their desired outcomes based on human annotation (Weller et al., 2020, Mishra et al., 2022, 2022, 2022, 2022, 2022, 2022, 2022). Additionally, they show a direct correlation between the size and diversity of the "instructional" data and the generalizability of resulting models to unseen tasks (Wang et al., 2022, 2022). However, since these developments largely focus on existing NLP tasks and depend on human-annotated instructions, this poses a bottleneck for progress toward more generalizable models (e.g., see Fig. 5a in Wang et al., 2022). Our work aims to move beyond classical NLP tasks and tackle the challenges of creating diverse instruction data by employing pretrained LMs. InstructGPT (Ouyang et al., 2022) shares a similar goal as ours in building more general-purpose LMs, and has demonstrated remarkable performance in following diverse user instructions. However, as a commercial system, their construction process still remains quite opaque. In particular, the role of _data_ has remained understudied due to limited transparency and the private user data they used in their study. Addressing such challenges necessitates the creation of a large-scale, public dataset covering a broad range of tasks.
๋ช…๋ น ํ›„์† LMs.A ์ผ๋ จ์˜ ์ž‘์—…์€ ๋ฐ”๋‹๋ผ LMs๊ฐ€ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ "๋ช…๋ น" ๋ฐ์ดํ„ฐ--์–ธ์–ด ๋ช…๋ น ๋ช…๋ น ๋ฐ ์ธ๊ฐ„ ์ฃผ์„์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ์›ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํŠœ๋‹๋˜๋ฉด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ช…๋ น์„ ๋”ฐ๋ฅด๋Š” ๋ฐ ํšจ๊ณผ์ ์ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ๋‹ค(Weller et al., 2020, Mishra et al., 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022). ๋˜ํ•œ, "๊ต์œก" ๋ฐ์ดํ„ฐ์˜ ํฌ๊ธฐ์™€ ๋‹ค์–‘์„ฑ๊ณผ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ ์‚ฌ์ด์— ์ง์ ‘์ ์ธ ์ƒ๊ด€ ๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ค€๋‹ค(Wang et al., 2022, 2022). ๊ทธ๋Ÿฌ๋‚˜, ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ๋“ค์€ ์ฃผ๋กœ ๊ธฐ์กด์˜ NLP ํƒœ์Šคํฌ๋“ค์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์ธ๊ฐ„ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ ๋ช…๋ น๋“ค์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ์ด๋Š” ๋ณด๋‹ค ์ผ๋ฐ˜ํ™”๋  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ๋“ค์„ ํ–ฅํ•œ ์ง„๋ณด์— ๋ณ‘๋ชฉ ํ˜„์ƒ์„ ์•ผ๊ธฐํ•œ๋‹ค(์˜ˆ๋ฅผ ๋“ค์–ด, Wang ๋“ฑ์˜ ๋„๋ฉด 5a ์ฐธ์กฐ, 2022). ์šฐ๋ฆฌ์˜ ์ž‘์—…์€ ๊ณ ์ „์ ์ธ NLP ์ž‘์—…์„ ๋„˜์–ด ์‚ฌ์ „ ํ›ˆ๋ จ๋œ LMs๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ˆ˜์—… ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“œ๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค. InstructGPT (Ouyang et al., 2022)๋Š” ๋ณด๋‹ค ๋ฒ”์šฉ์ ์ธ LMs๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์šฐ๋ฆฌ์™€ ์œ ์‚ฌํ•œ ๋ชฉํ‘œ๋ฅผ ๊ณต์œ ํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ์ž ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋Š” ๋ฐ ์žˆ์–ด ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ์—… ์‹œ์Šคํ…œ์œผ๋กœ์„œ ๊ทธ๋“ค์˜ ๊ฑด์„ค ๊ณผ์ •์€ ์—ฌ์ „ํžˆ ์ƒ๋‹นํžˆ ๋ถˆํˆฌ๋ช…ํ•˜๋‹ค. ํŠนํžˆ _data_์˜ ์—ญํ• ์€ ์ œํ•œ๋œ ํˆฌ๋ช…์„ฑ๊ณผ ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ๊ฐœ์ธ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ์•„์ง ์—ฐ๊ตฌ๋˜์ง€ ์•Š์€ ์ƒํƒœ๋กœ ๋‚จ์•„ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์„ ํฌํ•จํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Compared to the cost of training a language model on this dataset, the additional work required to deduplicate the training dataset is negligible.
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋น„์šฉ๊ณผ ๋น„๊ตํ•˜์—ฌ, ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์ถ”๊ฐ€ ์ž‘์—…์€ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋‹ค.
3. Replace \(x\) with \(-x\) in the expression: \(f(f(g(f(g(f(-x))))))\). Since \(f\) is odd, we have \(f(-x)=-f(x)\), so the expression becomes \(-f(f(g(f(g(-x)))))\).
3. ์‹์—์„œ \(f(f(g(f(g(f(-x))))))\๋กœ \(x\)๋ฅผ \(-x\)๋กœ ๋ฐ”๊ฟ‰๋‹ˆ๋‹ค. \(f\)๋Š” ํ™€์ˆ˜์ด๋ฏ€๋กœ ์šฐ๋ฆฌ๋Š” \(f(-x)=-f(x)\๋ฅผ ๊ฐ€์ง€๋ฏ€๋กœ ์‹์€ \(-f(f(g(f(g(-x)))))\๊ฐ€ ๋œ๋‹ค.
Multiple choiceIn this setup, models are given a prompt and then must select an answer from a predefined set [35; 36; 37]. For instance, in the BBQ dataset [38], models are given a statement like "A muslim and a hindu walk into a bar. Someone planted a bomb and everyone ran out". The model is then asked a question about this situation. BBQ includes both non-negative questions (e.g., "Who ran for cover?") and negative questions (e.g., "Who likely planted the bomb?"). Models' biases can be assessed by comparing the distribution of their multiple choice answers. Multiple choice tests can be scored automatically, are easy to interpret, and allow for fine-grained assessment of models. However, it can also be unrealistic given that most interactions in real-world applications do not involve multiple-choice questions. Further, recent studies have highlighted methodological limitations, demonstrating that altering the order of multiple-choice options [39; 40], rephrasing the question [40], or changing the multiple-choice options [41] can significantly increase the variability of the responses of the models.
๋‹ค์ค‘ ์„ ํƒ ์ด ์„ค์ •์—์„œ ๋ชจ๋ธ์—๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์ œ๊ณต๋œ ๋‹ค์Œ ๋ฏธ๋ฆฌ ์ •์˜๋œ ์„ธํŠธ[35; 36; 37]์—์„œ ๋‹ต๋ณ€์„ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, BBQ ๋ฐ์ดํ„ฐ ์„ธํŠธ [38]์—์„œ ๋ชจ๋ธ์—๋Š” "์ด์Šฌ๋žŒ๊ต๋„์™€ ํžŒ๋‘๊ฐ€ ์ˆ ์ง‘์— ๋“ค์–ด๊ฐ‘๋‹ˆ๋‹ค. ๋ˆ„๊ตฐ๊ฐ€ ํญํƒ„์„ ์‹ฌ์—ˆ๊ณ  ๋ชจ๋“  ์‚ฌ๋žŒ์ด ๋›ฐ์ณ๋‚˜๊ฐ”์Šต๋‹ˆ๋‹ค."์™€ ๊ฐ™์€ ์ง„์ˆ ์ด ์ฃผ์–ด์ง‘๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์€ ์ด ์ƒํ™ฉ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. BBQ๋Š” ๋ถ€์ •์ ์ด์ง€ ์•Š์€ ์งˆ๋ฌธ๋“ค(์˜ˆ๋ฅผ ๋“ค์–ด, "๋ˆ„๊ฐ€ ๋ฎ์น˜๊ธฐ ์œ„ํ•ด ๋›ฐ์—ˆ๋Š”๊ฐ€?")๊ณผ ๋ถ€์ •์ ์ธ ์งˆ๋ฌธ๋“ค(์˜ˆ๋ฅผ ๋“ค์–ด, "๋ˆ„๊ฐ€ ํญํƒ„์„ ์‹ฌ์—ˆ์„๊นŒ?")์„ ๋ชจ๋‘ ํฌํ•จํ•œ๋‹ค. ๋ชจ๋ธ๋“ค์˜ ํŽธํ–ฅ์„ ๊ฐ๊ด€์‹ ๋‹ต๋ณ€์˜ ๋ถ„ํฌ๋ฅผ ๋น„๊ตํ•˜์—ฌ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ฐ๊ด€์‹ ํ…Œ์ŠคํŠธ๋Š” ์ž๋™์œผ๋กœ ์ ์ˆ˜๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ๊ณ , ํ•ด์„ํ•˜๊ธฐ ์‰ฌ์šฐ๋ฉฐ, ๋ชจ๋ธ์˜ ์„ธ๋ฐ€ํ•œ ํ‰๊ฐ€๊ฐ€ ๊ฐ€๋Šฅํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์—์„œ ๋Œ€๋ถ€๋ถ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ๊ฐ๊ด€์‹ ์งˆ๋ฌธ์„ ํฌํ•จํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์„ ๊ฐ์•ˆํ•  ๋•Œ ๋น„ํ˜„์‹ค์ ์ผ ์ˆ˜๋„ ์žˆ๋‹ค. ๋˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ๊ด€์‹ ์˜ต์…˜[39; 40]์˜ ์ˆœ์„œ๋ฅผ ๋ณ€๊ฒฝํ•˜๊ฑฐ๋‚˜ ์งˆ๋ฌธ์„ ๋‹ค์‹œ ํ‘œํ˜„ํ•˜๊ฑฐ๋‚˜ ๊ฐ๊ด€์‹ ์˜ต์…˜[41]์„ ๋ณ€๊ฒฝํ•˜๋ฉด ๋ชจ๋ธ์˜ ์‘๋‹ต ๋ณ€๋™์„ฑ์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋ฐฉ๋ฒ•๋ก ์  ํ•œ๊ณ„๋ฅผ ๊ฐ•์กฐํ–ˆ๋‹ค.
Recently, the Mamba network has been widely exploited in medical image segmentation, as illustrated in Fig. 12 (e, f, g). For example, the Swin-Transformer based model Swin-UNet [223] attains 89.33/99.57/88.46 (Dice, IoU, Accuracy) on the MRI Cardiac dataset. In contrast, the Mamba-based UNet achieves comparable or even better segment results, such as the Mamba-UNet [67], Semi-Mamba-UNet [70], and Weak-Mamba-UNet [72]. These results fully demonstrate the effectiveness of Mamba architecture for medical image segmentation.
์ตœ๊ทผ Mamba ๋„คํŠธ์›Œํฌ๋Š” ๊ทธ๋ฆผ๊ณผ ๊ฐ™์ด ์˜๋ฃŒ ์˜์ƒ ๋ถ„ํ• ์— ๋„๋ฆฌ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ๋„ 12(e, f, g)์— ๋„์‹œ๋œ ๋ฐ”์™€ ๊ฐ™๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Swin-Transformer ๊ธฐ๋ฐ˜ ๋ชจ๋ธ Swin-UNet [223]์€ MRI Cardiac ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 89.33/99.57/88.46(Dice, IoU, Accuracy)์— ๋„๋‹ฌํ•œ๋‹ค. ๋Œ€์กฐ์ ์œผ๋กœ, Mamba-๊ธฐ๋ฐ˜ UNet์€ Mamba-UNet[67], Semi-Mamba-UNet[70], ๋ฐ Weak-Mamba-UNet[72]์™€ ๊ฐ™์€ ๋น„๊ต๊ฐ€๋Šฅํ•˜๊ฑฐ๋‚˜ ์‹ฌ์ง€์–ด ๋” ๋‚˜์€ ์„ธ๊ทธ๋จผํŠธ ๊ฒฐ๊ณผ๋“ค์„ ๋‹ฌ์„ฑํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์˜๋ฃŒ ์˜์ƒ ๋ถ„ํ• ์„ ์œ„ํ•œ Mamba ์•„ํ‚คํ…์ฒ˜์˜ ์œ ํšจ์„ฑ์„ ์ถฉ๋ถ„ํžˆ ์ž…์ฆํ•œ๋‹ค.
Doubanis a social network and database that allows users to create content related to literature and artistic works such as films, books, TV series, music, etc. We sample data from books, movies, and TV series, extracting metadata that includes ratings, detailed information on actors/crew, and long reviews. Then, we design three tasks in total: synopsis generation, review generation, and recommendations. For each task, we manually design various prompt templates and used these templates in combination with metadata to construct instructions. For synopsis generation and review generation, we construct instructions using prompt templates combined with movie or TV series names, with responses generated by Douban users. Then we remove responses with lengths shorter than a threshold and delete personal information and irrelevant content(e.g., "Subscribe our Official Accounts"). Additionally, we manually adjusted some instructions to add more complex implicit intents, aligning better with the details of the response.
๋‘๋ฐ”๋‹ˆ์Šค๋Š” ์˜ํ™”, ์ฑ…, TV ์‹œ๋ฆฌ์ฆˆ, ์Œ์•… ๋“ฑ๊ณผ ๊ฐ™์€ ๋ฌธํ•™ ๋ฐ ์˜ˆ์ˆ  ์ž‘ํ’ˆ๊ณผ ๊ด€๋ จ๋œ ์ฝ˜ํ…์ธ ๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ์†Œ์…œ ๋„คํŠธ์›Œํฌ ๋ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์ด๋‹ค. ์šฐ๋ฆฌ๋Š” ์ฑ…, ์˜ํ™”, TV ์‹œ๋ฆฌ์ฆˆ์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ‰์ , ๋ฐฐ์šฐ/์Šน๋ฌด์›์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด, ๊ธด ๋ฆฌ๋ทฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์‹œ๋†‰์‹œ์Šค ์ƒ์„ฑ, ๋ฆฌ๋ทฐ ์ƒ์„ฑ, ์ถ”์ฒœ์˜ ์„ธ ๊ฐ€์ง€ ์ž‘์—…์„ ์„ค๊ณ„ํ•œ๋‹ค. ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ์ˆ˜๋™์œผ๋กœ ์„ค๊ณ„ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ํ…œํ”Œ๋ฆฟ์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์™€ ๊ฒฐํ•ฉํ•˜์—ฌ ๋ช…๋ น์–ด๋ฅผ ๊ตฌ์„ฑํ–ˆ๋‹ค. ์‹œ๋†‰์‹œ์Šค ์ƒ์„ฑ ๋ฐ ๋ฆฌ๋ทฐ ์ƒ์„ฑ์„ ์œ„ํ•ด ์˜ํ™” ๋˜๋Š” TV ์‹œ๋ฆฌ์ฆˆ ์ด๋ฆ„๊ณผ ๊ฒฐํ•ฉ๋œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ๊ณผ ๋”๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒ์„ฑํ•œ ์‘๋‹ต์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์นจ์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ž„๊ณ„๊ฐ’๋ณด๋‹ค ์งง์€ ๊ธธ์ด์˜ ์‘๋‹ต์„ ์ œ๊ฑฐํ•˜๊ณ  ๊ฐœ์ธ ์ •๋ณด ๋ฐ ๊ด€๋ จ ์—†๋Š” ๋‚ด์šฉ(์˜ˆ: "๊ณต์‹ ๊ณ„์ • ๊ตฌ๋…")์„ ์‚ญ์ œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‘๋‹ต์˜ ์„ธ๋ถ€ ์ •๋ณด์™€ ๋” ์ž˜ ์ผ์น˜ํ•˜๋„๋ก ๋ณต์žกํ•œ ์•”์‹œ์  ์˜๋„๋ฅผ ์ถ”๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์ง€์นจ์„ ์ˆ˜๋™์œผ๋กœ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
In this section, we provide appropriate background and methodology as it relates to continual pre-training in the context of LLMs.
์ด ์ ˆ์—์„œ๋Š” LLM์˜ ๋งฅ๋ฝ์—์„œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ๊ณผ ๊ด€๋ จํ•˜์—ฌ ์ ์ ˆํ•œ ๋ฐฐ๊ฒฝ๊ณผ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ๊ณตํ•œ๋‹ค.
**Formatting Synthetic Data.** To reduce the burden of human annotation or manual collection, several semi-automated approaches [143] have been proposed for constructing instances by feeding existing instances into LLMs to synthesize diverse task descriptions and instances. As illustrated in Figure 11(c), the Self-Instruct method only needs 175 instances as the initial task pool. Then, they randomly select a few instances from the pool as demonstrations and prompt a LLM to generate new instructions and corresponding input-output pairs. After the quality and diversity filtering, newly generated instances would be added into the task pool. Hence, the synthetic method is an effective and economical way to generate large-scale instruction data for LLMs. However, the instances generated by the Self-Instruct method might be simplistic or lack the diversity. To improve the quality of synthetic int ructions, WizardLM [346] introduces Evol-Instruct by proposing in-depth and in-breadth evolving to enrich the complexity and diversity of the instances. Furthermore, Self-Align [347] establishes multiple human-aligned principles to filter the synthesized instances. It then employs these instances to train a LLM in order to yield more aligned instances. To enhance the quality of the instance output, researchers directly adopt human-written texts as the output and synthesize corresponding instructions using ICL examples [348].
**ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ˜•์‹ํ™”** ์ธ๊ฐ„ ์ฃผ์„ ๋˜๋Š” ์ˆ˜๋™ ์ปฌ๋ ‰์…˜์˜ ๋ถ€๋‹ด์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ธ์Šคํ„ด์Šค๋ฅผ LLM์— ๊ณต๊ธ‰ ํ•˜ ์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—… ์„ค๋ช… ๋ฐ ์ธ์Šคํ„ด์Šค๋ฅผ ํ•ฉ์„ฑ ํ•˜ ์—ฌ ์ธ์Šคํ„ด์Šค๋ฅผ ๊ตฌ์„ฑ ํ•˜๋Š” ์—ฌ๋Ÿฌ ๋ฐ˜์ž๋™ ์ ‘๊ทผ๋ฒ• [143]์ด ์ œ์•ˆ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋„ 11(c)์— ์˜ˆ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, Self-Instruct ๋ฐฉ๋ฒ•์€ ์ดˆ๊ธฐ ํƒœ์Šคํฌ ํ’€๋กœ์„œ 175๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋“ค๋งŒ์„ ํ•„์š”๋กœ ํ•œ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ’€์—์„œ ๋ช‡ ๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์‹œ๋ฒ”์œผ๋กœ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜๊ณ  LLM์„ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ƒˆ ๋ช…๋ น์–ด์™€ ํ•ด๋‹น ์ž…๋ ฅ ์ถœ๋ ฅ ์Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ’ˆ์งˆ ๋ฐ ๋‹ค์–‘์„ฑ ํ•„ํ„ฐ๋ง ํ›„ ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๊ฐ€ ์ž‘์—… ํ’€์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ํ•ฉ์„ฑ ๋ฐฉ๋ฒ•์€ LLMs์— ๋Œ€ํ•œ ๋Œ€๊ทœ๋ชจ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š”๋ฐ ํšจ๊ณผ์ ์ด๊ณ  ๊ฒฝ์ œ์ ์ธ ๋ฐฉ๋ฒ•์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Self-Instruct ๋ฐฉ๋ฒ•์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋Š” ๋‹จ์ˆœํ•˜๊ฑฐ๋‚˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ๋‹ค. ํ•ฉ์„ฑ int ructions์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด, WizardLM [346]์€ ์ธ์Šคํ„ด์Šค์˜ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ์ง„ํ™”ํ•˜๋Š” ์‹ฌ์ธต์ ์ด๊ณ  ๊ด‘๋ฒ”์œ„ํ•œ ๊ฒƒ์„ ์ œ์•ˆํ•จ์œผ๋กœ์จ Evol-Instruct๋ฅผ ๋„์ž…ํ•œ๋‹ค. ๋˜ํ•œ, Self-Align[347]์€ ํ•ฉ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋“ค์„ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์ˆ˜์˜ ์ธ๊ฐ„-์ •๋ ฌ ์›๋ฆฌ๋“ค์„ ํ™•๋ฆฝํ•œ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์ธ์Šคํ„ด์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ์ •๋ ฌ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ธ์Šคํ„ด์Šค ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด, ์—ฐ๊ตฌ์ž๋“ค์€ ์ง์ ‘ ์ธ๊ฐ„-์ž‘์„ฑ๋œ ํ…์ŠคํŠธ๋“ค์„ ์ถœ๋ ฅ์œผ๋กœ ์ฑ„ํƒํ•˜๊ณ  ICL ์˜ˆ๋“ค์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€์‘ํ•˜๋Š” ๋ช…๋ น์–ด๋“ค์„ ํ•ฉ์„ฑํ•œ๋‹ค[348].
Given EntRec, we answer RQ1 by measuring how often the internal recall of \(e_{2}\) improves at layer \(l\) when modifying a two-hop prompt from \(\tau_{\text{2H}}^{\prime}\) to \(\tau_{\text{2H}}\), where \(\tau_{\text{2H}}^{\prime}\) does not contain the descriptive mention of \(e_{2}\) while \(\tau_{\text{2H}}\) does. To be specific, we measure _the relative frequency_ of \(\tau_{\text{2H}}\) in TwoHopFact where EntRec\({}^{l}(e_{2},\tau_{\text{2H}})>\)EntRec\({}^{l}(e_{2},\tau_{\text{2H}}^{\prime})\).
EntRec๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์šฐ๋ฆฌ๋Š” 2-ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ \(\tau_{\text{2H}}^{\prime}\)์—์„œ \(\tau_{\text{2H}}\)๋กœ ์ˆ˜์ •ํ•  ๋•Œ \(e_{2}\)์˜ ๋‚ด๋ถ€ ํšŒ์ƒ์ด ์ธต \(l\)์—์„œ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ๊ฐœ์„ ๋˜๋Š”์ง€๋ฅผ ์ธก์ •ํ•˜์—ฌ RQ1์— ์‘๋‹ตํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ \(\tau_{\text{2H}}^{\prime}\)์€ \(e_{2}\)์— ๋Œ€ํ•œ ์„ค๋ช…์  ์–ธ๊ธ‰์„ ํฌํ•จํ•˜์ง€ ์•Š๋Š” ๋ฐ˜๋ฉด \(\tau_{\text{2H}}\)๋Š” ํฌํ•จํ•œ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})>\)EntRec\({}^{l}(e_{2},\tau_{\text{2H}}^{\prime})\)์—์„œ \(\tau_{\text{2H}}\)์˜ ์ƒ๋Œ€์  ๋นˆ๋„์ˆ˜๋ฅผ ์ธก์ •ํ•œ๋‹ค.
Fig. 14: Flan-PaLM finetuning consist of 473 datasets in above task categories. Courtesy of [74].
๋„. 14: Flan-PaLM ๋ฏธ์„ธ ์กฐ์ •์€ ์œ„์˜ ์ž‘์—… ๋ฒ”์ฃผ์—์„œ 473๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. [74]์˜ ์˜ˆ์ ˆ.
Hoffmann et al.'s paper has been highly influential in the language modeling community. Its findings have informed the scaling policies of notable models, such as Google's Gemini suite (Google, 2023), and have likely guided the development of many other models. Moreover, the specific parameter estimates of the scaling law presented in the paper are of significant scientific interest, such as for machine learning theory. Given the wide-reaching impact of this research, it is crucial to thoroughly investigate the robustness and reproducibility of the work.
Hoffmann ๋“ฑ์˜ ๋…ผ๋ฌธ์€ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์˜ํ–ฅ๋ ฅ์ด ๋งค์šฐ ์ปธ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ๊ตฌ๊ธ€์˜ ์ œ๋ฏธ๋‹ˆ ์Šค์œ„ํŠธ๋ฃธ(๊ตฌ๊ธ€, 2023)๊ณผ ๊ฐ™์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋ชจ๋ธ์˜ ์Šค์ผ€์ผ๋ง ์ •์ฑ…์— ์•Œ๋ ค์กŒ๊ณ , ๋‹ค๋ฅธ ๋งŽ์€ ๋ชจ๋ธ์˜ ๊ฐœ๋ฐœ์„ ์ด๋Œ์—ˆ์„ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋‹ค. ๋”์šฑ์ด, ๋ณธ ๋…ผ๋ฌธ์—์„œ ์ œ์‹œ๋œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ํŠน์ • ๋งค๊ฐœ๋ณ€์ˆ˜ ์ถ”์ •์น˜๋Š” ๊ธฐ๊ณ„ ํ•™์Šต ์ด๋ก ๊ณผ ๊ฐ™์€ ์ค‘์š”ํ•œ ๊ณผํ•™์  ๊ด€์‹ฌ์‚ฌ์ด๋‹ค. ์ด ์—ฐ๊ตฌ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์˜ํ–ฅ์„ ๊ฐ์•ˆํ•  ๋•Œ ์ž‘์—…์˜ ๊ฒฌ๊ณ ์„ฑ๊ณผ ์žฌํ˜„์„ฑ์„ ์ฒ ์ €ํžˆ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค.
For each instruction \(I_{i}\), we extract the corresponding use case \(u_{i}\) and set of skills \(\mathbf{s}_{i}\). We then have the set of metadata as \(\mathcal{M}=\{(u_{i},\mathbf{s}_{i})\}_{i=1}^{n}\). Instructions may share or partially overlap in their \(u_{i}\)'s and \(\mathbf{s}_{i}\), reflecting the distribution of tasks and capabilities within the seed instructions. Use cases and skills are generated on-the-fly, not limited to some predefined sets, enabling broader applicability. However, we can always provide such constraints with our prior knowledge, or even directly write out metadata without any seed instructions.
๊ฐ ๋ช…๋ น์–ด \(I_{i}\)์— ๋Œ€ํ•ด ํ•ด๋‹น ์œ ์Šค์ผ€์ด์Šค \(u_{i}\)์™€ ์Šคํ‚ฌ ์ง‘ํ•ฉ \(\mathbf{s}_{i}\)์„ ์ถ”์ถœํ•œ๋‹ค. ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์€ \(\mathcal{M}=\{(u_{i},\mathbf{s}_{i})\}_{i=1}^{n}\)์ด๋‹ค. ๋ช…๋ น์–ด๋Š” \(u_{i}\)'s ๋ฐ \(\mathbf{s}_{i}\)์—์„œ ๊ณต์œ ๋˜๊ฑฐ๋‚˜ ๋ถ€๋ถ„์ ์œผ๋กœ ์ค‘์ฒฉ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์‹œ๋“œ ๋ช…๋ น์–ด๋“ค ๋‚ด์˜ ํƒœ์Šคํฌ๋“ค ๋ฐ ๋Šฅ๋ ฅ๋“ค์˜ ๋ถ„ํฌ๋ฅผ ๋ฐ˜์˜ํ•œ๋‹ค. ์œ ์Šค ์ผ€์ด์Šค ๋ฐ ์Šคํ‚ฌ์€ ์ผ๋ถ€ ๋ฏธ๋ฆฌ ์ •์˜๋œ ์„ธํŠธ์— ๊ตญํ•œ๋˜์ง€ ์•Š๊ณ  ์ฆ‰์‹œ ์ƒ์„ฑ๋˜์–ด ๋” ๋„“์€ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์šฐ๋ฆฌ๋Š” ํ•ญ์ƒ ์‚ฌ์ „ ์ง€์‹์œผ๋กœ ๊ทธ๋Ÿฌํ•œ ์ œ์•ฝ๋“ค์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๊ฑฐ๋‚˜, ์‹ฌ์ง€์–ด ์–ด๋– ํ•œ ์‹œ๋“œ ๋ช…๋ น๋„ ์—†์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ง์ ‘ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.
Back-translation augmentation.Many multilingual language models use back-translation as a data augmentation method, creating synthetic parallel training data from monolingual data sources (Bi et al., 2021; Caswell et al., 2019; Liao et al., 2021; Marie et al., 2020; Pham et al., 2021; Sennrich et al., 2016; Xu et al., 2022; Zheng et al., 2020). For example, Sennrich et al. (2016) back-translate monolingual target data into source language data, providing additional parallel training samples for substantial translation task improvements. Researchers have also explored different sampling methods for back-translation (e.g., beam search, constrained sampling, unconstrained sampling) and their comparative effectiveness (Edunov et al., 2018; Graca et al., 2019; Sennrich et al., 2016). Xu et al. (2022) emphasize the importance of the weight and quality of synthetic data for optimal NMT performance using back-translation. They propose a method to optimize the ratio between search methods and a gamma score to balance estimated importance weight and quality. However, some limitations exist with back-translation-based synthetic data generation. For example, the quality and diversity of synthetic data depends on the performance of the back-translation method. If the synthetic data is too noisy or not diverse, the performance gain would be limited (Chauhan et al., 2022; Epaliyana et al., 2021).
์—ญ-๋ฒˆ์—ญ ์ฆ๊ฐ•.๋งŽ์€ ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ๋“ค์€ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์œผ๋กœ์„œ ์—ญ-๋ฒˆ์—ญ์„ ์‚ฌ์šฉํ•˜์—ฌ, ๋‹จ์ผ ์–ธ์–ด ๋ฐ์ดํ„ฐ ์†Œ์Šค๋“ค๋กœ๋ถ€ํ„ฐ ํ•ฉ์„ฑ ๋ณ‘๋ ฌ ํŠธ๋ ˆ์ด๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ๋‹ค(Bi et al., 2021; Caswell et al., 2019; Liao et al., 2021; Marie et al., 2020; Pham et al., 2021; Sennrich et al., 2016; Xu et al., 2022; Zheng et al., 2020). ์˜ˆ๋ฅผ ๋“ค์–ด, Sennrich ๋“ฑ(2016)์€ ๋‹จ์ผ ์–ธ์–ด ํƒ€๊ฒŸ ๋ฐ์ดํ„ฐ๋ฅผ ์†Œ์Šค ์–ธ์–ด ๋ฐ์ดํ„ฐ๋กœ ์—ญ๋ฒˆ์—ญํ•˜์—ฌ, ์‹ค์งˆ์ ์ธ ๋ฒˆ์—ญ ์ž‘์—… ๊ฐœ์„ ์„ ์œ„ํ•œ ์ถ”๊ฐ€์ ์ธ ๋ณ‘๋ ฌ ํŠธ๋ ˆ์ด๋‹ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•œ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ๋˜ํ•œ ์—ญ-๋ฒˆ์—ญ(์˜ˆ๋ฅผ ๋“ค์–ด, ๋น” ํƒ์ƒ‰, ์ œ์•ฝ๋œ ์ƒ˜ํ”Œ๋ง, ์ œ์•ฝ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ๋ง)์„ ์œ„ํ•œ ์ƒ์ดํ•œ ์ƒ˜ํ”Œ๋ง ๋ฐฉ๋ฒ• ๋ฐ ์ด๋“ค์˜ ๋น„๊ต ํšจ๊ณผ๋ฅผ ํƒ์ƒ‰ํ•˜์˜€๋‹ค(Edunov et al., 2018; Graca et al., 2019; Sennrich et al., 2016). Xu et al.(2022)์€ ์—ญ๋ณ€ํ™˜์„ ์ด์šฉํ•œ ์ตœ์ ์˜ NMT ์„ฑ๋Šฅ์„ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์ค‘์น˜์™€ ํ’ˆ์งˆ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•œ๋‹ค. ๊ทธ๋“ค์€ ์ถ”์ •๋œ ์ค‘์š”๋„ ๊ฐ€์ค‘์น˜์™€ ํ’ˆ์งˆ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๊ฒ€์ƒ‰ ๋ฐฉ๋ฒ•๊ณผ ๊ฐ๋งˆ ์ ์ˆ˜ ์‚ฌ์ด์˜ ๋น„์œจ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—ญ๋ฒˆ์—ญ ๊ธฐ๋ฐ˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ œํ•œ ์‚ฌํ•ญ์ด ์žˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์€ ์—ญ๋ฒˆ์—ญ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋„ˆ๋ฌด ์žก์Œ์ด ๋งŽ๊ฑฐ๋‚˜ ๋‹ค์–‘ํ•˜์ง€ ์•Š๋‹ค๋ฉด, ์„ฑ๋Šฅ ์ด๋“์€ ์ œํ•œ๋  ๊ฒƒ์ด๋‹ค(Chauhan et al., 2022; Epaliyana et al., 2021).
requiring extensive reasoning and deep understanding, larger models are more suitable; whereas for situations with limited resources or lower task complexity, smaller models may be a more economical choice.
๊ด‘๋ฒ”์œ„ํ•œ ์ถ”๋ก ๊ณผ ๊นŠ์€ ์ดํ•ด๋ฅผ ํ•„์š”๋กœ ํ•˜๋Š” ๊ฒฝ์šฐ ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ ํ•ฉํ•˜์ง€๋งŒ ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ์ ์ด๊ฑฐ๋‚˜ ์ž‘์—… ๋ณต์žก๋„๊ฐ€ ๋‚ฎ์€ ์ƒํ™ฉ์—์„œ๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ์ด ๋” ๊ฒฝ์ œ์ ์ธ ์„ ํƒ์ด ๋  ์ˆ˜ ์žˆ๋‹ค.
To sum up, we initiate an investigation into quantitative prediction methods for pretraining data curation. With its increasing focus on data engineering, we hope our exploration facilitates further quantitative studies and theoretical analyses in this research area.
์š”์•ฝํ•˜๋ฉด, ์šฐ๋ฆฌ๋Š” ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์„ ์œ„ํ•œ ์ •๋Ÿ‰์  ์˜ˆ์ธก ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์กฐ์‚ฌ๋ฅผ ์‹œ์ž‘ํ•œ๋‹ค. ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง์— ๋”ฐ๋ผ ๋ณธ ์—ฐ๊ตฌ์˜ ํƒ์ƒ‰์ด ์ด ์—ฐ๊ตฌ ์˜์—ญ์—์„œ ์ถ”๊ฐ€ ์–‘์  ์—ฐ๊ตฌ์™€ ์ด๋ก ์  ๋ถ„์„์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ๋ฅผ ๋ฐ”๋ž€๋‹ค.
Nemotron-4 uses a standard decoder-only Transformer architecture (Vaswani et al., 2017), with causal attention masks. Exact hyper-parameters affecting size are shown in Table 1. Nemotron-4 has 3.2 billion embedding parameters and 12.5 billion non-embedding parameters. We use Rotary Position Embeddings (RoPE) (Su et al., 2021), SentencePiece tokenizer (Kudo and Richardson, 2018), squared ReLU activations in the MLP layers, no bias terms, dropout rate of zero, and untied input-output embeddings. We use grouped
๋„ค๋ชจํŠธ๋ก -4๋Š” ์ธ๊ณผ์  ์ฃผ์˜ ๋งˆ์Šคํฌ์™€ ํ•จ๊ป˜ ํ‘œ์ค€ ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜(Vaswani et al., 2017)๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ํฌ๊ธฐ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์ •ํ™•ํ•œ ํ•˜์ดํผ-ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ‘œ 1์— ๋‚˜ํƒ€๋‚˜ ์žˆ๋‹ค. ๋„ค๋ชจํŠธ๋ก -4๋Š” 32์–ต ๊ฐœ์˜ ์ž„๋ฒ ๋”ฉ ํŒŒ๋ผ๋ฏธํ„ฐ์™€ 125์–ต ๊ฐœ์˜ ๋น„- ์ž„๋ฒ ๋”ฉ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ–๋Š”๋‹ค. Rotary Position Embeddings (RoPE) (Su et al., 2021), SentencePiece tokenizer (Kudo and Richardson, 2018), ์ œ๊ณฑ ReLU activations in the MLP layers, no bias terms, dropout rate of zero, and unied input-output embeddings. ๊ทธ๋ฃนํ™” ์‚ฌ์šฉ
It is noteworthy that the relative performance gain is more substantial with increasing task difficulty. Specifically, we observe that within the same task, the smaller model, Llama2-13B, gains ranging from \(28\%\)-\(200\%\), but only \(8\%\)-\(16\%\) over GPT-3.5-Turbo. Moreover, the more challenging task MATH yields gains of \(34\%\)-\(200\%\), in contrast to only \(16\%\)-\(69\%\) on the easier task GSM8K.
์ฃผ๋ชฉํ•  ์ ์€ ๊ณผ์ œ ๋‚œ์ด๋„๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์ƒ๋Œ€์  ์„ฑ๊ณผ ์ด๋“์ด ๋” ํฌ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ํŠนํžˆ, ๋™์ผํ•œ ์ž‘์—… ๋‚ด์—์„œ ๋” ์ž‘์€ ๋ชจ๋ธ์ธ Llama2-13B๋Š” GPT-3.5-Turbo๋ณด๋‹ค \(28\%\)-\(200\%\)์˜ ์ด๋“ ๋ฒ”์œ„๋ฅผ ๊ฐ–์ง€๋งŒ \(8\%\)-\(16\%\)์— ๋ถˆ๊ณผํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•œ๋‹ค. ๋”์šฑ์ด, ๋” ์–ด๋ ค์šด ํƒœ์Šคํฌ MATH๋Š” ๋” ์‰ฌ์šด ํƒœ์Šคํฌ GSM8K์—์„œ ๋‹จ์ง€ \(16\%\)-\(69\%\)์— ๋น„ํ•ด \(34\%\)-\(200\%\)์˜ ์ด๋“์„ ์–ป๋Š”๋‹ค.
Manual analysis of the GPT-4 predictions.The authors of this paper manually assess randomly sampled 20 instances for each aspect and check if GPT-4 predictions match their assessments given the same instruction, demonstrations, and test instances. We found our assessments show high agreement with GPT-4 predictions, especially for relevance (95%), retrieval necessity (95%), and the degree of support (90%). Agreement was slightly lower in usefulness (80%), mostly due to the disagreement between 1 and 2 or 4 and 5.
GPT-4 ์˜ˆ์ธก์˜ ์ˆ˜๋™ ๋ถ„์„.๋ณธ ๋…ผ๋ฌธ์˜ ์ €์ž๋Š” ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ 20๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ๋™์ผํ•œ ๋ช…๋ น, ์‹œ์—ฐ ๋ฐ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์ฃผ์–ด์ง„ ๊ฒฝ์šฐ GPT-4 ์˜ˆ์ธก์ด ํ‰๊ฐ€์™€ ์ผ์น˜ํ•˜๋Š”์ง€ ํ™•์ธํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ํ‰๊ฐ€๊ฐ€ ํŠนํžˆ ๊ด€๋ จ์„ฑ(95%), ๊ฒ€์ƒ‰ ํ•„์š”์„ฑ(95%), ์ง€์› ์ •๋„(90%)์— ๋Œ€ํ•ด GPT-4 ์˜ˆ์ธก๊ณผ ๋†’์€ ์ผ์น˜๋ฅผ ๋ณด์ธ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค. ๋™์˜๋Š” ์œ ์šฉ์„ฑ(80%)์—์„œ ์•ฝ๊ฐ„ ๋‚ฎ์•˜๋Š”๋ฐ, ์ด๋Š” ๋Œ€๋ถ€๋ถ„ 1๊ณผ 2 ๋˜๋Š” 4์™€ 5 ์‚ฌ์ด์˜ ๋ถˆ์ผ์น˜๋กœ ์ธํ•œ ๊ฒƒ์ด์—ˆ๋‹ค.
**Important Findings from Existing Work**. Recently, a very comprehensive evaluation [421] has been conducted about the impact of multiple factors (_e.g.,_ model size and sensitivity) on the post-training quantization methods. Another study [422] examines the scaling law of \(k\)-bit quantization in inference performance. In addition to the overall performance, the study [423] specifically focuses on the potential impact of quantification on emergent capabilities, as well as the levels of performance that can be achieved across various levels of bit precision. Also, prior work (_e.g.,_ LLM.int8(424), GPTQ [417], QLoRA [419], and GLM [93]) has also extensively examined the performance of quantization methods in various settings. Next, we summarize several important findings from these studies, which will be useful for those who may not want to delve into the technical details of quantization methods.
**๊ธฐ์กด ์ž‘์—…์˜ ์ค‘์š”ํ•œ ๊ฒฐ๊ณผ**. ์ตœ๊ทผ, ๋‹ค์ค‘ ์ธ์ž(์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋ธ ํฌ๊ธฐ ๋ฐ ๋ฏผ๊ฐ๋„)๊ฐ€ ํ›ˆ๋ จ ํ›„ ์–‘์žํ™” ๋ฐฉ๋ฒ•์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•ด ๋งค์šฐ ํฌ๊ด„์ ์ธ ํ‰๊ฐ€[421]๊ฐ€ ์ˆ˜ํ–‰๋˜์—ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ[422]๋Š” ์ถ”๋ก  ์„ฑ๋Šฅ์—์„œ \(k\)-๋น„ํŠธ ์–‘์žํ™”์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์กฐ์‚ฌํ•œ๋‹ค. ์ „์ฒด ์„ฑ๋Šฅ ์™ธ์—๋„ ์—ฐ๊ตฌ[423]์€ ํŠนํžˆ ๋‹ค์–‘ํ•œ ์ˆ˜์ค€์˜ ๋น„ํŠธ ์ •๋ฐ€๋„์— ๊ฑธ์ณ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์„ฑ๋Šฅ ์ˆ˜์ค€๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ฐฝ๋ฐœ ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ์ •๋Ÿ‰ํ™”์˜ ์ž ์žฌ์  ์˜ํ–ฅ์— ์ดˆ์ ์„ ๋งž์ถ˜๋‹ค. ๋˜ํ•œ ์ด์ „ ์ž‘์—…(_์˜ˆ:_ LLM.int8(424), GPTQ[417], QLoRA[419], GLM[93])๋„ ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ ์–‘์žํ™” ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์กฐ์‚ฌํ–ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์–‘์žํ™” ๋ฐฉ๋ฒ•์˜ ๊ธฐ์ˆ ์  ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํŒŒ๊ณ ๋“ค์ง€ ์•Š์„ ์ˆ˜ ์žˆ๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ์œ ์šฉํ•  ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์˜ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์š”์•ฝํ•œ๋‹ค.
_This summary is largely factual and looks at the overall arc of the text, though it has some chronology problems (Ar's proposal comes after the run in the city, which comes after Josh and Radhyu open the pub-up), and focuses disproportionately on the author's access, despite their complex emotional journey. The author's correction with another writer, Kate Robb, as opposed to the text itself. It also doesn't discuss the nonlinear storyellings._
_์ด ์š”์•ฝ์€ ๋Œ€์ฒด๋กœ ์‚ฌ์‹ค์ ์ด๋ฉฐ ํ…์ŠคํŠธ์˜ ์ „๋ฐ˜์ ์ธ ํ˜ธ๋ฅผ ์‚ดํŽด๋ณด์ง€๋งŒ ๋ช‡ ๊ฐ€์ง€ ์—ฐ๋Œ€๊ธฐ ๋ฌธ์ œ๊ฐ€ ์žˆ๋‹ค(Ar์˜ ์ œ์•ˆ์€ ์กฐ์‰ฌ์™€ ๋ผ๋“€๊ฐ€ ํŽ์—…์„ ์—ฐ ํ›„์— ๋‚˜์˜ค๋Š” ๋„์‹œ ์šด์˜ ํ›„์— ๋‚˜์˜ค๋ฉฐ ๋ณต์žกํ•œ ๊ฐ์ • ์—ฌ์ •์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ €์ž์˜ ์ ‘๊ทผ์— ๋ถˆ๊ท ํ˜•์ ์œผ๋กœ ์ดˆ์ ์„ ๋งž์ถ˜๋‹ค). ์ž‘๊ฐ€๊ฐ€ ํ…์ŠคํŠธ ๊ทธ ์ž์ฒด์™€๋Š” ๋ฐ˜๋Œ€๋กœ ๋‹ค๋ฅธ ์ž‘๊ฐ€ ์ผ€์ดํŠธ ๋กญ๊ณผ ๊ต์ •ํ•œ ๊ฒƒ์ด๋‹ค. ๋˜ํ•œ ๋น„์„ ํ˜• ์Šคํ† ๋ฆฌํ…”๋ง์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค._
Figure 21: Examples on instruction following. MiniCPM-2.4B-DPO successfully follows the instruction and generates a poem with correct tonal pattern and the first character of each sentence. However, MiniCPM-1.2B only generates the correct tonal pattern.
๋„ 21: ๋‹ค์Œ์˜ ์ง€์‹œ์— ๋Œ€ํ•œ ์˜ˆ. MiniCPM-2.4B-DPO๋Š” ๋ช…๋ น์„ ์„ฑ๊ณต์ ์œผ๋กœ ๋”ฐ๋ฅด๊ณ , ์ •ํ™•ํ•œ ์„ฑ์กฐ ํŒจํ„ด๊ณผ ๊ฐ ๋ฌธ์žฅ์˜ ์ฒซ ๋ฒˆ์งธ ๋ฌธ์ž๋ฅผ ๊ฐ€์ง„ ์‹œ๋ฅผ ์ƒ์„ฑํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ MiniCPM-1.2B๋Š” ์˜ฌ๋ฐ”๋ฅธ ์„ฑ์กฐ ํŒจํ„ด๋งŒ์„ ์ƒ์„ฑํ•œ๋‹ค.
From the results presented in Table V it is clear that GPT-4 achieves best results for HellaSwag while Davinci-003 is best model for OBQA. It is also good to note that results for OBQA are not reported for all of the models and possibly davinci-003 is not the best model achieving highest results on OBQA.
ํ‘œ V์— ์ œ์‹œ๋œ ๊ฒฐ๊ณผ์—์„œ GPT-4๊ฐ€ ํ—ฌ๋ผ์Šค์›จ๊ทธ์— ๋Œ€ํ•ด ์ตœ์ƒ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ˜๋ฉด ๋‹ค๋นˆ์น˜-003์€ OBQA์— ๋Œ€ํ•ด ์ตœ์ƒ์˜ ๋ชจ๋ธ์ž„์ด ๋ถ„๋ช…ํ•˜๋‹ค. OBQA์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๋ชจ๋“  ๋ชจ๋ธ์— ๋Œ€ํ•ด ๋ณด๊ณ ๋˜์ง€ ์•Š์œผ๋ฉฐ ๋‹ค๋นˆ์น˜-003์ด OBQA์—์„œ ๊ฐ€์žฅ ๋†’์€ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ์ตœ์ƒ์˜ ๋ชจ๋ธ์ด ์•„๋‹ˆ๋ผ๋Š” ์ ์— ์ฃผ๋ชฉํ•˜๋Š” ๊ฒƒ๋„ ์ข‹๋‹ค.
1. Rephrased and Expanded Question: Is it accurate to state that the film 'Barbie' will achieve a domestic box office gross that is at least double the amount that 'Oppenheimer' will earn during their respective opening weekends? This comparison will be based on the final reported figures for the three-day opening weekend (July 21-23) in the United States, as documented on Box Office Mojo. The outcome will be determined by comparing the "Domestic Weekend" grosses of 'Barbie' and 'Oppenheimer,' and the market will resolve in favor of 'Yes' if 'Barbie's gross is at least twice that of 'Oppenheimer.' If no final data is available by July 31, 2023, an alternative credible source will be used to resolve the market. 2. Reasons why the answer might be no:
1. ์š”์•ฝ ๋ฐ ํ™•์žฅ ์งˆ๋ฌธ: ์˜ํ™” '๋ฐ”๋น„'๊ฐ€ ๊ฐœ๋ด‰ ์ฃผ๋ง ๋™์•ˆ '์˜คํŽœํ•˜์ด๋จธ'๊ฐ€ ๋ฒŒ์–ด๋“ค์ผ ๊ธˆ์•ก์˜ ์ตœ์†Œ ๋‘ ๋ฐฐ์ธ ๊ตญ๋‚ด ํฅํ–‰ ์ด์•ก์„ ๋‹ฌ์„ฑํ•  ๊ฒƒ์ด๋ผ๊ณ  ๋งํ•˜๋Š” ๊ฒƒ์ด ์ •ํ™•ํ•œ๊ฐ€? ์ด ๋น„๊ต๋Š” ๋ฐ•์Šค์˜คํ”ผ์Šค ๋ชจ์กฐ์— ๊ธฐ๋ก๋œ ๋ฐ”์™€ ๊ฐ™์ด ๋ฏธ๊ตญ์—์„œ 3์ผ๊ฐ„์˜ ๊ฐœ๋ง‰ ์ฃผ๋ง(7์›” 21์ผ-23์ผ)์— ๋Œ€ํ•œ ์ตœ์ข… ๋ณด๊ณ ๋œ ์ˆ˜์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•  ๊ฒƒ์ด๋‹ค. ๊ฒฐ๊ณผ๋Š” '๋ฐ”๋น„'์™€ '์˜คํŽœํ•˜์ด๋จธ'์˜ "๊ตญ๋‚ด ์ฃผ๋ง" ์ด์ˆ˜์ต์„ ๋น„๊ตํ•˜์—ฌ ๊ฒฐ์ •๋˜๋ฉฐ, '๋ฐ”๋น„์˜ ์ด์ˆ˜์ต์ด '์˜คํŽœํ•˜์ด๋จธ'์˜ ์ตœ์†Œ 2๋ฐฐ๋ผ๋ฉด ์‹œ์žฅ์€ '์˜ˆ'์— ์ฐฌ์„ฑํ•˜์—ฌ ํ•ด๊ฒฐ๋œ๋‹ค. 2023๋…„ 7์›” 31์ผ๊นŒ์ง€ ์ตœ์ข… ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋ฉด ์‹œ์žฅ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€์ฒด ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•  ๊ฒƒ์ด๋‹ค. 2. ๋Œ€๋‹ต์ด '์•„๋‹ˆ์˜ค'์ผ ์ˆ˜ ์žˆ๋Š” ์ด์œ :
making them publicly available, so they are referred to as close-source models. In our research, we extensively integrate a significant number of close-source models as the foundational benchmarks. Specifically, our baselines encompass the following models: (i) OpenAI's GPT-3 [51], GPT-3.5, ChatGPT5, GPT-4 [3]; (ii) Google's PaLM 2 [44], PaLM [7], and Minerva [15]; (iii) Anthropic's Claude Instant [39], Claude 1.3\({}^{\circ}\), Claude 2\({}^{\prime}\), DeepMind's Chinchilla [47].
๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ทผ์ ‘ ์†Œ์Šค ๋ชจ๋ธ์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ์—์„œ ์šฐ๋ฆฌ๋Š” ๊ธฐ์ดˆ ๋ฒค์น˜๋งˆํฌ๋กœ ์ƒ๋‹นํ•œ ์ˆ˜์˜ ๊ทผ์ ‘ ์†Œ์Šค ๋ชจ๋ธ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ํ†ตํ•ฉํ•œ๋‹ค. ํŠนํžˆ, OpenAI์˜ GPT-3[51], GPT-3.5, ChatGPT5, GPT-4[3], ๊ตฌ๊ธ€์˜ PaLM2[44], PaLM[7], ๋ฏธ๋„ค๋ฅด๋ฐ”[15], (iii) Anthropic์˜ Claude Instant[39], Claude 1.3\({}^{\circ}\), Claude 2\({}^{\prime}\), DeepMind์˜ Chinchilla[47].
Figure 10: Confusion matrix comparing human annotations to model predictions for the claim verification task. The left side shows the counts for predictions made by Claude-3-Opus contrasted with human labels. The right side shows the count for predictions made by GPT-4-Turbo contrasted with human labels. A more detailed breakdown is presented in Table 32.
๋„ 10: ์ฒญ๊ตฌ ๊ฒ€์ฆ ํƒœ์Šคํฌ์— ๋Œ€ํ•œ ๋ชจ๋ธ ์˜ˆ์ธก๋“ค์— ๋Œ€ํ•œ ์ธ๊ฐ„ ์ฃผ์„๋“ค์„ ๋น„๊ตํ•˜๋Š” ํ˜ผ๋™ ํ–‰๋ ฌ. ์™ผ์ชฝ์€ Claude-3-Opus๊ฐ€ ์ธ๊ฐ„ ๋ผ๋ฒจ๊ณผ ๋Œ€์กฐํ•˜์—ฌ ๋งŒ๋“  ์˜ˆ์ธก์˜ ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ์˜ค๋ฅธ์ชฝ์€ ์ธ๊ฐ„ ๋ผ๋ฒจ๊ณผ ๋Œ€์กฐ๋˜๋Š” GPT-4-ํ„ฐ๋ณด์˜ ์˜ˆ์ธก ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๋ณด๋‹ค ์ž์„ธํ•œ ์„ค๋ช…์€ ํ‘œ 32์— ๋‚˜์™€ ์žˆ๋‹ค.
**Normalization Methods.** Training instability is a challenging issue for pre-training LLMs. To alleviate this issue, normalization is a widely adopted strategy to stabilize the training of neural networks. In the vanilla Transformer [22], LayerNorm [256] is employed. Recently, several advanced normalization techniques have been proposed as alternatives to LayerNorm, _e.g._, RMSNorm, and DeepNorm.
**์ •๊ทœํ™” ๋ฐฉ๋ฒ•** ํ›ˆ๋ จ ๋ถˆ์•ˆ์ •์„ฑ์€ ์‚ฌ์ „ ํ›ˆ๋ จ LLM์— ๋Œ€ํ•ด ์–ด๋ ค์šด ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ •๊ทœํ™”๋Š” ์‹ ๊ฒฝ๋ง์˜ ํ•™์Šต์„ ์•ˆ์ •ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋„๋ฆฌ ์ฑ„ํƒ๋˜๋Š” ์ „๋žต์ด๋‹ค. ๋ฐ”๋‹๋ผ ํŠธ๋žœ์Šคํฌ๋จธ[22]์—๋Š” LayerNorm[256]์ด ์ฑ„์šฉ๋œ๋‹ค. ์ตœ๊ทผ์—๋Š” LayerNorm, _e.g._, RMSNorm ๋ฐ DeepNorm์˜ ๋Œ€์•ˆ์œผ๋กœ ๋ช‡ ๊ฐ€์ง€ ๊ณ ๊ธ‰ ์ •๊ทœํ™” ๊ธฐ์ˆ ์ด ์ œ์•ˆ๋˜์—ˆ๋‹ค.
Fig. 39: HuggingGPT: An agent-based approach to use tools and planning [image courtesy of [171]]* **RACE**[186] suits for reading comprehension task. This dataset is based on English tests completed by Chinese students from middle school and high school, aged \(12\) to \(18\), and it contains roughly \(28,000\) texts and \(100,000\) questions rigorously prepared by human specialists, primarily English instructors. This dataset contains a wide range of subjects that were purposefully chosen to assess students' comprehension and reasoning abilities. This dataset is available in three subgroups: RACE-M, RACE-H, and RACE. RACE-M refers to the middle school examinations, whereas RACE-H denotes the high school tests. Finally, RACE is the synthesis of RACE-M and RACE-H. * **SQuAD**[187] stands for "Stanford Question Answering Dataset" and is a crowdsourced reading comprehension dataset based on Wikipedia articles. It has approximately \(100,000\) question-answer pairs connected to more than \(500\) articles. The answers to these questions are typically text fragments or spans taken from the corresponding reading passages. The questions may be unanswerable in some cases. The dataset is divided into three sets: an \(80\%\) training set, a \(10\%\) development set, and a \(10\%\) hidden test set.
๋„. 39: HuggingGPT: ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์ฝ๊ธฐ ์ดํ•ด ์ž‘์—…์„ ์œ„ํ•œ [์ด๋ฏธ์ง€ ์˜ˆ์ ˆ [171]]* **RACE**[186] ์ŠˆํŠธ๋ฅผ ๊ณ„ํšํ•˜๋Š” ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘ํ•™๊ต์™€ ๊ณ ๋“ฑํ•™๊ต์˜ ์ค‘๊ตญ ํ•™์ƒ๋“ค์ด ์ž‘์„ฑํ•œ ์˜์–ด ํ…Œ์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ฃผ๋กœ ์˜์–ด ๊ฐ•์‚ฌ์ธ ์ธ๊ฐ„ ์ „๋ฌธ๊ฐ€์— ์˜ํ•ด ์—„๊ฒฉํ•˜๊ฒŒ ์ค€๋น„๋œ ๋Œ€๋žต์ ์ธ \(28,000\) ํ…์ŠคํŠธ์™€ \(100,000\) ์งˆ๋ฌธ์„ ํฌํ•จํ•œ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ํ•™์ƒ๋“ค์˜ ์ดํ•ด๋ ฅ๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์˜๋„์ ์œผ๋กœ ์„ ํƒ๋œ ๊ด‘๋ฒ”์œ„ํ•œ ๊ณผ๋ชฉ์ด ํฌํ•จ๋˜์–ด ์žˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” RACE-M, RACE-H ๋ฐ RACE์˜ ์„ธ ํ•˜์œ„ ๊ทธ๋ฃน์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. RACE-M์€ ์ค‘ํ•™๊ต ์‹œํ—˜์„ ์˜๋ฏธํ•˜๋Š” ๋ฐ˜๋ฉด, RACE-H๋Š” ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜์„ ์˜๋ฏธํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, RACE๋Š” RACE-M๊ณผ RACE-H์˜ ํ•ฉ์„ฑ์ด๋‹ค. * **SQuAD**[187]์€ "Stanford Question Answering Dataset"์˜ ์•ฝ์ž์ด๋ฉฐ ์œ„ํ‚คํ”ผ๋””์•„ ๊ธฐ์‚ฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ํฌ๋ผ์šฐ๋“œ์†Œ์‹ฑ ์ฝ๊ธฐ ์ดํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์•ฝ \(100,000\)์˜ ์งˆ์˜์‘๋‹ต ์Œ์ด \(500\) ์ด์ƒ์˜ ๋…ผ๋ฌธ์— ์—ฐ๊ฒฐ๋˜์–ด ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ๋“ค์— ๋Œ€ํ•œ ๋‹ต๋ณ€๋“ค์€ ์ „ํ˜•์ ์œผ๋กœ ๋Œ€์‘ํ•˜๋Š” ํŒ๋… ํŒจ์‹œ์ง€๋“ค๋กœ๋ถ€ํ„ฐ ์ทจํ•ด์ง„ ํ…์ŠคํŠธ ๋‹จํŽธ๋“ค ๋˜๋Š” ์ŠคํŒฌ๋“ค์ด๋‹ค. ๊ทธ ์งˆ๋ฌธ๋“ค์€ ์–ด๋–ค ๊ฒฝ์šฐ์—๋Š” ๋Œ€๋‹ตํ•  ์ˆ˜ ์—†์„ ์ˆ˜๋„ ์žˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” \(80\%\) ํ›ˆ๋ จ ์„ธํŠธ, \(10\%\) ๊ฐœ๋ฐœ ์„ธํŠธ ๋ฐ \(10\%\) ์ˆจ๊ฒจ์ง„ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์„ธ ์„ธ ์„ธ ์„ธํŠธ๋กœ ๋‚˜๋‰œ๋‹ค.
Our VLM's enhanced performance on the JA-VG-VQA-500 benchmark indicates its proficiency in Japanese, highlighting the successful integration of the source Japanese LLM with the LLM component of the original VLM through evolutionary merging. Furthermore, our VLM's superior results on the JA-VLM-Bench-In-the-Wild compared to both baselines exhibits its adeptness at navigating culturally-specific content.
JA-VG-VQA-500 ๋ฒค์น˜๋งˆํฌ์—์„œ VLM์ด ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์€ ์ผ๋ณธ์–ด์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ง„ํ™”์  ๋ณ‘ํ•ฉ์„ ํ†ตํ•ด ์›๋ณธ VLM์˜ LLM ๊ตฌ์„ฑ ์š”์†Œ์™€ ์†Œ์Šค ์ผ๋ณธ LLM์ด ์„ฑ๊ณต์ ์œผ๋กœ ํ†ตํ•ฉ๋˜์—ˆ์Œ์„ ๊ฐ•์กฐํ•œ๋‹ค. ๋˜ํ•œ, ๋‘ ๊ธฐ์ค€์„ ์— ๋น„ํ•ด JA-VLM-Bench-In-Wild์— ๋Œ€ํ•œ VLM์˜ ์šฐ์ˆ˜ํ•œ ๊ฒฐ๊ณผ๋Š” ๋ฌธํ™”์ ์œผ๋กœ ํŠน์ •ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํƒ์ƒ‰ํ•˜๋Š” ๋ฐ ๋Šฅ์ˆ™ํ•จ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.
To measure the impact of our trained tokenizers on the model downstream performance, we trained one model for each tokenizer. In particular, for each of our 24 trained tokenizers, we trained a 2.6B transformer-based decoder-only model on up to 52B tokens following the scaling law proposed by (Hoffmann et al., 2022). Additionally, serving as baselines, we trained a monolingual and a multilingual model using the pre-trained GPT-2 tokenizer (Radford et al., 2018). All models have been trained based on the causal language modeling training objective.
ํ›ˆ๋ จ๋œ ํ† ํฐ๋ผ์ด์ €๊ฐ€ ๋ชจ๋ธ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ํ† ํฐ๋ผ์ด์ €์— ๋Œ€ํ•ด ํ•˜๋‚˜์˜ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ๋‹ค. ํŠนํžˆ, 24๊ฐœ์˜ ํ›ˆ๋ จ๋œ ํ† ํฐ์•„์ด์ € ๊ฐ๊ฐ์— ๋Œ€ํ•ด, (Hoffmann et al., 2022)์— ์˜ํ•ด ์ œ์•ˆ๋œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ์ตœ๋Œ€ 52B ํ† ํฐ์— ๋Œ€ํ•ด 2.6B ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œ์ผฐ๋‹ค. ์ถ”๊ฐ€์ ์œผ๋กœ, ๋ฒ ์ด์Šค๋ผ์ธ์œผ๋กœ์„œ, ์šฐ๋ฆฌ๋Š” ์‚ฌ์ „ ํŠธ๋ ˆ์ด๋‹๋œ GPT-2 ํ† ๋‚˜์ด์ €๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ์–ธ์–ด ๋ฐ ๋‹ค์ค‘ ์–ธ์–ด ๋ชจ๋ธ์„ ํŠธ๋ ˆ์ด๋‹ํ•˜์˜€๋‹ค(Radford et al., 2018). ๋ชจ๋“  ๋ชจ๋ธ์€ ์ธ๊ณผ ์–ธ์–ด ๋ชจ๋ธ๋ง ํ›ˆ๋ จ ๋ชฉํ‘œ์— ๊ธฐ์ดˆํ•˜์—ฌ ํ›ˆ๋ จ๋˜์—ˆ๋‹ค.
While chain-of-thought reasoning has showcased remarkable performance on numerous tasks, some challenges still require further exploration. In this section, we provide a concise overview of three promising avenues for future research: multi-modal X-of-thought reasoning (SS6.1), faithful X-of-thought reasoning (SS6.2), and X-of-thought reasoning theory (SS6.3).
์—ฐ์‡„์  ์‚ฌ๊ณ  ์ถ”๋ก ์€ ์ˆ˜๋งŽ์€ ๊ณผ์ œ์—์„œ ๋†€๋ผ์šด ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ ์ผ๋ถ€ ๊ณผ์ œ๋Š” ์—ฌ์ „ํžˆ ์ถ”๊ฐ€ ํƒ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ์ด ์ ˆ์—์„œ๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์„ธ ๊ฐ€์ง€ ์œ ๋งํ•œ ๋ฐฉ๋ฒ•, ์ฆ‰ ๋‹ค์ค‘ ๋ชจ๋“œ X-์‚ฌ์ƒ ์ถ”๋ก (SS6.1), ์ถฉ์‹คํ•œ X-์‚ฌ์ƒ ์ถ”๋ก (SS6.2), X-์‚ฌ์ƒ ์ถ”๋ก  ์ด๋ก (SS6.3)์— ๋Œ€ํ•œ ๊ฐ„๋žตํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•œ๋‹ค.
\(\bullet\)_Prompt learning with sufficient data._ In this approach, most existing methods regard continuous prompts as trainable model parameters and then leverage supervised learning to optimize the continuous prompts by minimizing the cross-entropy loss based on sufficient downstream task data [475, 396, 397, 401]. As discussed in Section 5.3.1, prefix tuning [396] prepends a sequence of prefixes (_i.e.,_ a set of trainable continuous vectors) to each Transformer layer in language models, while prompt tuning [397] only incorporates trainable prompt vectors at the input layer. By fixing the large-scale parameters of LLMs and only tuning continuous prompt vector, this kind of approaches can be extremely parameter-efficient (Section 5.3). However, these approaches are typically independent of the inputs, lacking sufficient consideration of input semantics. Therefore, the authors in [475] propose context tuning, where the continuous prompts are derived based on the input text and learned through the downstream task losses.
\(\bullet\)_์ถฉ๋ถ„ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ–๋Š” ํ”„๋กฌํ”„ํŠธ ํ•™์Šต._ ์ด ์ ‘๊ทผ๋ฒ•์—์„œ, ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ๋ฐฉ๋ฒ•๋“ค์€ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ›ˆ๋ จ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ฐ„์ฃผํ•˜๊ณ , ์ด์–ด์„œ ์ถฉ๋ถ„ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ํƒœ์Šคํฌ ๋ฐ์ดํ„ฐ์— ๊ธฐ์ดˆํ•˜์—ฌ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค์„ ์ตœ์†Œํ™”ํ•จ์œผ๋กœ์จ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ง€๋„ ํ•™์Šต์„ ํ™œ์šฉํ•œ๋‹ค[475, 396, 397, 401]. ์„น์…˜ 5.3.1์—์„œ ๋…ผ์˜๋œ ๋ฐ”์™€ ๊ฐ™์ด, ํ”„๋ฆฌํ”ฝ์Šค ํŠœ๋‹[396]์€ ํ”„๋ฆฌํ”ฝ์Šค๋“ค์˜ ์‹œํ€€์Šค(_์ฆ‰, ํ›ˆ๋ จ๊ฐ€๋Šฅํ•œ ์—ฐ์† ๋ฒกํ„ฐ๋“ค์˜ ์„ธํŠธ)๋ฅผ ์–ธ์–ด ๋ชจ๋ธ๋“ค์—์„œ ๊ฐ๊ฐ์˜ ํŠธ๋žœ์Šคํฌ๋จธ ๊ณ„์ธต์œผ๋กœ ํ”„๋ฆฌํŽœ์Šคํ•˜๋Š” ๋ฐ˜๋ฉด, ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹[397]์€ ์ž…๋ ฅ ๊ณ„์ธต์—์„œ ํ›ˆ๋ จ๊ฐ€๋Šฅํ•œ ํ”„๋กฌํ”„ํŠธ ๋ฒกํ„ฐ๋“ค๋งŒ์„ ํ†ตํ•ฉํ•œ๋‹ค. LLM์˜ ๋Œ€๊ทœ๋ชจ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ณ ์ •ํ•˜๊ณ  ์—ฐ์† ํ”„๋กฌํ”„ํŠธ ๋ฒกํ„ฐ๋งŒ ํŠœ๋‹ํ•จ์œผ๋กœ์จ, ์ด๋Ÿฌํ•œ ์ข…๋ฅ˜์˜ ์ ‘๊ทผ๋ฒ•์€ ๋งค์šฐ ๋งค๊ฐœ๋ณ€์ˆ˜ ํšจ์œจ์ ์ผ ์ˆ˜ ์žˆ๋‹ค(์„น์…˜ 5.3). ๊ทธ๋Ÿฌ๋‚˜, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ๋ฒ•๋“ค์€ ํ†ต์ƒ์ ์œผ๋กœ ์ž…๋ ฅ ์˜๋ฏธ๋“ค์— ๋Œ€ํ•œ ์ถฉ๋ถ„ํ•œ ๊ณ ๋ ค๊ฐ€ ๊ฒฐ์—ฌ๋œ ์ž…๋ ฅ๋“ค๊ณผ ๋…๋ฆฝ์ ์ด๋‹ค. ๋”ฐ๋ผ์„œ [475]์˜ ์ €์ž๋“ค์€ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์œ ๋„๋˜๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ํƒœ์Šคํฌ ์†์‹ค์„ ํ†ตํ•ด ํ•™์Šต๋˜๋Š” ์ปจํ…์ŠคํŠธ ํŠœ๋‹์„ ์ œ์•ˆํ•œ๋‹ค.
* _Fine-grained quantization_. For Transformer models, weights and activations are usually represented in the form of tensors. A straightforward approach is to use coarse-grained quantization parameters for the whole tensor (_i.e.,_ per-tensor quantization) [414]. However, it usually leads to inaccurate reconstruction results. Thus, fine-grained methods are proposed to reduce the quantization error. ZeroQuant [415] adopts a token-wise quantization approach with dynamic calibration for compressing activations. Whereas for weights (easier to be quantized), it uses a group-wise quantization. In practice, a group size of 128 [415, 416] is commonly used for model quantization. * _Balancing the quantization difficulty_. Considering that weights are easier to be quantized than activations, SmoothQuant [414] proposes to migrate the difficulty from activations to weights. Specially, they incorporate a scaling transformation to balance the difficulty between weights and activations in a linear layer: \(\mathbf{Y}=(\mathbf{X}\text{diag}(\mathbf{s})^{-1})\cdot(\text{diag}(\mathbf{ s})\mathbf{W})\). By introducing an mathematically equivalent transformation, this formula controls the quantization difficulty through the scaling factor \(\mathbf{s}\). To set \(\mathbf{s}\), it incorporates a migration strength parameter \(\alpha\) to balance the difficulties, where each entry \(s_{j}=\max(\mathbf{x}_{j})^{\alpha}/\max(\mathbf{w}_{j})^{(1-\alpha)}\) is determined by the migration strength. * _Layerwise quantization_. This approach finds optimal quantized weights that minimize a layerwise reconstruction loss: \(\arg\min_{\widetilde{\mathbf{W}}}\parallel\mathbf{W}\mathbf{X}-\widetilde{ \mathbf{W}}\mathbf{X}\parallel_{2}^{2}\). To efficiently optimize this objective, GPTQ [417] improves the original optimal brain quantization (OBO) [418] method by fixing the quantization order of weights for all rows. Further, with specially designed methods (_i.e.,_ lazy batch-updates and Cholesky reformulation), GPTQ is feasible to quantize very large models (_e.g.,_ 175B OPT) in 3 or 4 bit precision. More recently, AWQ [416] further simplifies the optimization form by incorporating activation-aware scaling for weights, which resembles the idea of SmoothQuant [414]: weights corresponding to outlier activations are more important to be precisely quantized. It does not directly optimize the reconstruction loss, but instead performs simple hyper-parameter search to achieve the minimal loss on calibration data.
* _์„ธ๋ฆฝ ์–‘์žํ™”_. ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ€์ค‘์น˜ ๋ฐ ํ™œ์„ฑํ™”๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํ…์„œ์˜ ํ˜•ํƒœ๋กœ ํ‘œํ˜„๋œ๋‹ค. ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ๋ฒ•์€ ์ „์ฒด ํ…์„œ(_์ฆ‰,_ํ…์„œ๋‹น ์–‘์žํ™”)์— ๋Œ€ํ•ด ๊ฑฐ์นœ-๊ฒฐ์ •ํ™”๋œ ์–‘์žํ™” ํŒŒ๋ผ๋ฏธํ„ฐ๋“ค์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด๋‹ค[414]. ๊ทธ๋Ÿฌ๋‚˜, ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ€์ •ํ™•ํ•œ ์žฌ๊ตฌ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์ดˆ๋ž˜ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ์–‘์žํ™” ์˜ค์ฐจ๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ์„ธ๋ฐ€ํ•œ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ์ œ๋กœํ€€ํŠธ[415]๋Š” ํ™œ์„ฑํ™”๋ฅผ ์••์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ๋™์  ์บ˜๋ฆฌ๋ธŒ๋ ˆ์ด์…˜์„ ๊ฐ–๋Š” ํ† ํฐ-์™€์ด์ฆˆ ์–‘์žํ™” ์ ‘๊ทผ๋ฒ•์„ ์ฑ„ํƒํ•œ๋‹ค. ๊ฐ€์ค‘์น˜(์–‘์žํ™”๋˜๊ธฐ ์‰ฌ์›€)์— ๋Œ€ํ•ด์„œ๋Š” ๊ทธ๋ฃน ๋‹จ์œ„ ์–‘์žํ™”๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์‹ค์ œ๋กœ, 128[415, 416]์˜ ๊ทธ๋ฃน ํฌ๊ธฐ๊ฐ€ ๋ชจ๋ธ ์–‘์žํ™”๋ฅผ ์œ„ํ•ด ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋œ๋‹ค. * _์–‘์žํ™” ๋‚œ์ด๋„ ๊ท ํ˜• ์กฐ์ •_ ์ž…๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜๊ฐ€ ํ™œ์„ฑํ™”๋ณด๋‹ค ์–‘์žํ™”ํ•˜๊ธฐ ์‰ฝ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•˜์—ฌ SmoothQuant[414]๋Š” ํ™œ์„ฑํ™”์—์„œ ๊ฐ€์ค‘์น˜๋กœ ๋‚œ์ด๋„๋ฅผ ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•œ๋‹ค. ํŠนํžˆ \(\mathbf{Y}=(\mathbf{X}\text{diag}(\mathbf{s})^{-1})\cdot(\text{diag}(\mathbf{ s})\mathbf{W})\). ์ˆ˜ํ•™์ ์œผ๋กœ ๋™๋“ฑํ•œ ๋ณ€ํ™˜์„ ๋„์ž…ํ•จ์œผ๋กœ์จ, ์ด ๊ณต์‹์€ ์Šค์ผ€์ผ๋ง ์ธ์ž \(\mathbf{s}\)๋ฅผ ํ†ตํ•ด ์–‘์žํ™” ๋‚œ์ด๋„๋ฅผ ์ œ์–ดํ•œ๋‹ค. \(\mathbf{s}\)์„ ์„ค์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ ๊ฐ•๋„ ๋งค๊ฐœ๋ณ€์ˆ˜ \(\alpha\)๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์–ด๋ ค์›€์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋ฉฐ, ์—ฌ๊ธฐ์„œ ๊ฐ ์—”ํŠธ๋ฆฌ \(s_{j}=\max(\mathbf{x}_{j})^{\alpha}/\max(\mathbf{w}_{j})^{(1-\alpha)}\)๋Š” ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ ๊ฐ•๋„์— ์˜ํ•ด ๊ฒฐ์ •๋œ๋‹ค. * _๊ณ„์ธต๋ณ„ ์–‘์žํ™”_. ์ด ๋ฐฉ๋ฒ•์€ ๊ณ„์ธต๋ณ„ ์žฌ๊ตฌ์„ฑ ์†์‹ค์„ ์ตœ์†Œํ™”ํ•˜๋Š” ์ตœ์ ์˜ ์–‘์žํ™” ๊ฐ€์ค‘์น˜๋ฅผ ์ฐพ๋Š”๋‹ค : \(\arg\min_{\widetilde{\mathbf{W}}}\parallel\mathbf{W}\mathbf{X}-\widetilde{ \mathbf{W}}\mathbf{X}\parallel_{2}^{2}\). ์ด ๋ชฉ์ ์„ ํšจ์œจ์ ์œผ๋กœ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด, GPTQ[417]์€ ๋ชจ๋“  ํ–‰์— ๋Œ€ํ•œ ๊ฐ€์ค‘์น˜์˜ ์–‘์žํ™” ์ˆœ์„œ๋ฅผ ๊ณ ์ •ํ•จ์œผ๋กœ์จ ์›๋ž˜์˜ ์ตœ์  ๋ธŒ๋ ˆ์ธ ์–‘์žํ™”(OBO)[418] ๋ฐฉ๋ฒ•์„ ๊ฐœ์„ ํ•œ๋‹ค. ๋˜ํ•œ, ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ๋ฐฉ๋ฒ•(_์ฆ‰,_ lazy batch-updates ๋ฐ Cholesky reformulation)์„ ์‚ฌ์šฉํ•˜์—ฌ GPTQ๋Š” 3 ๋˜๋Š” 4 ๋น„ํŠธ ์ •๋ฐ€๋„๋กœ ๋งค์šฐ ํฐ ๋ชจ๋ธ(_์˜ˆ๋ฅผ ๋“ค์–ด,_175B OPT)์„ ์–‘์žํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋ณด๋‹ค ์ตœ๊ทผ์—, AWQ[416]์€ ๊ฐ€์ค‘์น˜์— ๋Œ€ํ•œ ํ™œ์„ฑํ™”-์ธ์‹ ์Šค์ผ€์ผ๋ง์„ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ์ตœ์ ํ™” ํ˜•ํƒœ๋ฅผ ๋”์šฑ ๋‹จ์ˆœํ™”ํ•˜๋Š”๋ฐ, ์ด๋Š” SmoothQuant[414]์˜ ์•„์ด๋””์–ด์™€ ์œ ์‚ฌํ•˜๋‹ค : ์ด์ƒ์น˜ ํ™œ์„ฑํ™”์— ๋Œ€์‘ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋Š” ์ •๋ฐ€ํ•˜๊ฒŒ ์–‘์žํ™”๋˜๋Š” ๊ฒƒ์ด ๋” ์ค‘์š”ํ•˜๋‹ค. ์ด๊ฒƒ์€ ์žฌ๊ตฌ์„ฑ ์†์‹ค์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ, ๋Œ€์‹ ์— ๊ต์ • ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ตœ์†Œ ์†์‹ค์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ํ•˜์ดํผ-ํŒŒ๋ผ๋ฏธํ„ฐ ํƒ์ƒ‰์„ ์ˆ˜ํ–‰ํ•œ๋‹ค.
This chain-of-reasoning strategy enabled AMIE to progressively refine its response conditioned on the current conversation to arrive at an informed and grounded reply.
์ด ์ถ”๋ก  ์—ฐ์‡„ ์ „๋žต์„ ํ†ตํ•ด AMIE๋Š” ํ˜„์žฌ ๋Œ€ํ™”์— ๋Œ€ํ•ด ์กฐ๊ฑดํ™”๋œ ์‘๋‹ต์„ ์ ์ง„์ ์œผ๋กœ ์ •์ œํ•˜์—ฌ ์ •๋ณด์— ์ž…๊ฐํ•˜๊ณ  ๊ทผ๊ฑฐ ์žˆ๋Š” ์‘๋‹ต์— ๋„๋‹ฌํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค.
Optimization of Memory and Storage Architecture. In the context of agent collaboration and communication, the future design of memory and storage systems can adopt a shared approach, enabling the sharing of memory and storage between agents. Such an architecture would enable agents to access a communal pool of memory and storage, thereby improving the agents' decision-making ability since one agent can benefit from other agents' memory or storage. Moreover, future work can explore hierarchical storage solutions, designed to optimize data retrieval and storage efficiency. This could involve prioritizing quicker access and reduced storage allocation for frequently accessed data, and vice versa for less frequently accessed information.
๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ์•„ํ‚คํ…์ฒ˜์˜ ์ตœ์ ํ™” ์—์ด์ „ํŠธ ํ˜‘์—… ๋ฐ ํ†ต์‹ ์˜ ๋งฅ๋ฝ์—์„œ, ํ–ฅํ›„ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ์‹œ์Šคํ…œ์˜ ์„ค๊ณ„๋Š” ๊ณต์œ  ์ ‘๊ทผ๋ฒ•์„ ์ฑ„ํƒํ•˜์—ฌ, ์—์ด์ „ํŠธ๋“ค ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€์˜ ๊ณต์œ ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•„ํ‚คํ…์ฒ˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์ €์žฅ์†Œ์˜ ๊ณต๋™ ํ’€์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์—ฌ, ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋ฉ”๋ชจ๋ฆฌ ๋˜๋Š” ์ €์žฅ์†Œ์˜ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—์ด์ „ํŠธ์˜ ์˜์‚ฌ ๊ฒฐ์ • ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค. ๋˜ํ•œ, ํ–ฅํ›„ ์ž‘์—…์€ ๋ฐ์ดํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ํšจ์œจ์„ฑ์„ ์ตœ์ ํ™”ํ•˜๋„๋ก ์„ค๊ณ„๋œ ๊ณ„์ธต์  ์Šคํ† ๋ฆฌ์ง€ ์†”๋ฃจ์…˜์„ ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ์ž์ฃผ ์•ก์„ธ์Šค๋˜๋Š” ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋” ๋น ๋ฅธ ์•ก์„ธ์Šค ๋ฐ ๊ฐ์†Œ๋œ ์Šคํ† ๋ฆฌ์ง€ ํ• ๋‹น์˜ ์šฐ์„ ์ˆœ์œ„๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ๊ณ , ๋œ ์ž์ฃผ ์•ก์„ธ์Šค๋˜๋Š” ์ •๋ณด์— ๋Œ€ํ•ด์„œ๋Š” ๊ทธ ๋ฐ˜๋Œ€์˜ ๊ฒฝ์šฐ๋„ ๋งˆ์ฐฌ๊ฐ€์ง€์ด๋‹ค.
Xueyang Feng is currently studying for a Ph.D. degree at Renmin University of China, Beijing, China. His research interests include recommender system, agent based on large language model.
์‰์–‘ํŽ‘์€ ํ˜„์žฌ ์ค‘๊ตญ ๋ฒ ์ด์ง• ๋Ÿฐ๋ฏผ๋Œ€์—์„œ ๋ฐ•์‚ฌ ํ•™์œ„๋ฅผ ๋ฐ›๊ณ  ์žˆ๋‹ค. ๊ทธ์˜ ์—ฐ๊ตฌ ๊ด€์‹ฌ์‚ฌ๋กœ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ถ”์ฒœ ์‹œ์Šคํ…œ, ์—์ด์ „ํŠธ ๋“ฑ์ด ์žˆ๋‹ค.
In addition to human evaluations, we implemented model-based auto-evaluation methods as economical consistent alternatives to specialist assessments. These techniques were employed to evaluate both dialogue quality and diagnostic accuracy of the OSCE agent. To establish the validity of our auto-evaluation methods for assessing dialogue quality, we initially focused on a subset of four evaluation axes from the PACES rubric(Table A.2) that were assessed by both the patient actors and the specialist physicians. The auto-evaluation, which uses a self-CoT strategy (details described in Section A.9) with AMIE to rate dialogues, was in good alignment with human raters and comparable to the inter-specialist agreement on these criteria. For the auto-evaluation of differential diagnoses, we leveraged another LLM, Med-PaLM 2 [13] as a surrogate for a specialist rater to grade the predicted diagnoses against the ground truth diagnoses (more details in Section A.7). Our auto-evaluation on DDx accuracy showed a similar trend for AMIE and OSCE agents compared to the specialist ratings. Overall, auto-evaluation trends aligned with human ratings for both dialogue quality and diagnostic accuracy.
์ธ๊ฐ„ ํ‰๊ฐ€ ์™ธ์—๋„ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์— ๋Œ€ํ•œ ๊ฒฝ์ œ์ ์ธ ์ผ๊ด€๋œ ๋Œ€์•ˆ์œผ๋กœ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ๊ตฌํ˜„ํ–ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์€ OSCE ์—์ด์ „ํŠธ์˜ ๋Œ€ํ™” ํ’ˆ์งˆ๊ณผ ์ง„๋‹จ ์ •ํ™•๋„๋ฅผ ๋ชจ๋‘ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ๋Œ€ํ™” ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ž๋™ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์˜ ํƒ€๋‹น์„ฑ์„ ํ™•๋ฆฝํ•˜๊ธฐ ์œ„ํ•ด ์ฒ˜์Œ์—๋Š” ํ™˜์ž ํ–‰์œ„์ž์™€ ์ „๋ฌธ์˜ ์˜์‚ฌ ๋ชจ๋‘๊ฐ€ ํ‰๊ฐ€ํ•œ PACES ๋ฃจ๋ธŒ๋ฆญ(ํ‘œ A.2)์˜ 4๊ฐ€์ง€ ํ‰๊ฐ€ ์ถ• ํ•˜์œ„ ์ง‘ํ•ฉ์— ์ดˆ์ ์„ ๋งž์ท„๋‹ค. ๋Œ€ํ™”๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด AMIE์™€ ํ•จ๊ป˜ ์ž์ฒด CoT ์ „๋žต(์„น์…˜ A.9์— ์„ค๋ช…๋œ ์„ธ๋ถ€ ์‚ฌํ•ญ)์„ ์‚ฌ์šฉํ•˜๋Š” ์ž๋™ ํ‰๊ฐ€๋Š” ์ธ๊ฐ„ ํ‰๊ฐ€์ž์™€ ์ž˜ ์ผ์น˜ํ–ˆ์œผ๋ฉฐ ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€ ๊ฐ„ ํ•ฉ์˜์— ํ•„์ ํ–ˆ๋‹ค. ์ฐจ๋“ฑ ์ง„๋‹จ์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ๋˜ ๋‹ค๋ฅธ LLM, Med-PaLM 2 [13]์„ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์ž์˜ ๋Œ€๋ฆฌ์ธ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ์˜ˆ์ธก๋œ ์ง„๋‹จ์„ ๊ทธ๋ผ์šด๋“œ ํŠธ๋ฃจ์Šค ์ง„๋‹จ์— ๋Œ€ํ•ด ๋“ฑ๊ธ‰์„ ๋งค๊ฒผ๋‹ค(์„น์…˜ A.7์˜ ๋” ์ž์„ธํ•œ ๋‚ด์šฉ). DDx ์ •ํ™•๋„์— ๋Œ€ํ•œ ์šฐ๋ฆฌ์˜ ์ž๋™ ํ‰๊ฐ€๋Š” ์ „๋ฌธ๊ฐ€ ๋“ฑ๊ธ‰๊ณผ ๋น„๊ตํ•˜์—ฌ AMIE ๋ฐ OSCE ์—์ด์ „ํŠธ์— ๋Œ€ํ•ด ์œ ์‚ฌํ•œ ๊ฒฝํ–ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ์ž๋™ ํ‰๊ฐ€ ๊ฒฝํ–ฅ์€ ๋Œ€ํ™” ํ’ˆ์งˆ๊ณผ ์ง„๋‹จ ์ •ํ™•๋„ ๋ชจ๋‘์— ๋Œ€ํ•ด ์ธ๊ฐ„ ๋“ฑ๊ธ‰๊ณผ ์ผ์น˜ํ–ˆ๋‹ค.
Note how the loss scales linearly with the number of generations. Figure 3, middle, illustrates how an increased number
์†์‹ค์ด ์„ธ๋Œ€ ์ˆ˜์— ๋”ฐ๋ผ ์„ ํ˜•์ ์œผ๋กœ ํ™•์žฅ๋˜๋Š” ๋ฐฉ๋ฒ•์— ์œ ์˜ํ•˜์‹ญ์‹œ์˜ค. ๋„ 3, ์ค‘๊ฐ„, ์ฆ๊ฐ€๋œ ์ˆซ์ž๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๋ฐฉ๋ฒ•
In the context of LLMs, an agent refers to a system based on a specialized instantiation of an (augmented) LLM that is capable of performing specific tasks autonomously. These agents are designed to interact with users and environment to make decisions based on the input and the intended goal of the interaction. Agents are based on LLMs equipped with the
LLM์˜ ๋งฅ๋ฝ์—์„œ ์—์ด์ „ํŠธ๋Š” ํŠน์ • ์ž‘์—…์„ ์ž์œจ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” (์ฆ๊ฐ•๋œ) LL์˜ ํŠน์ˆ˜ ์ธ์Šคํ„ด์Šคํ™”์— ๊ธฐ๋ฐ˜ํ•œ ์‹œ์Šคํ…œ์„ ์˜๋ฏธํ•œ๋‹ค. ์ด๋“ค ์—์ด์ „ํŠธ๋Š” ์ž…๋ ฅ ๋ฐ ์ƒํ˜ธ์ž‘์šฉ์˜ ์˜๋„๋œ ๋ชฉํ‘œ์— ๊ธฐ์ดˆํ•˜์—ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ์ž ๋ฐ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ์ž‘์šฉํ•˜๋„๋ก ์„ค๊ณ„๋œ๋‹ค. ์—์ด์ „ํŠธ๋Š” ํƒ‘์žฌ๋œ LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Pretraining CorpusFor mathematical reasoning, we utilize the OpenWebMath (OWM) dataset (Paster et al., 2023), which comprises approximately 14B tokens sourced from math-related web pages in the Common Crawl. In the general domain, we combine the SlimPajama (Daria et al., 2023)
Pretraining Corpus ์ˆ˜ํ•™์  ์ถ”๋ก ์„ ์œ„ํ•ด Common Crawl์˜ ์ˆ˜ํ•™ ๊ด€๋ จ ์›น ํŽ˜์ด์ง€์—์„œ ๊ฐ€์ ธ์˜จ ์•ฝ 14B ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ OpenWebM(OWM) ๋ฐ์ดํ„ฐ ์„ธํŠธ(Paster ๋“ฑ, 2023)๋ฅผ ํ™œ์šฉํ•œ๋‹ค. ์ผ๋ฐ˜ ๋„๋ฉ”์ธ์—์„œ, ์šฐ๋ฆฌ๋Š” SlimPajama๋ฅผ ๊ฒฐํ•ฉํ•œ๋‹ค(Daria et al., 2023).
Weak-to-Strong GenerationApart from the main experiments where we use the same base model for the reference and continual pretraining, we also investigate if a smaller reference model can effectively guide the pretraining of a larger model. We use Tinyllma-1.1B as reference model and continual pretraining Llama-2-7B on math. Results presented in Table 3 indicate that, despite the considerable gap between the small and large models (Li et al., 2023), employing the small reference model to token selection can still yield benefits to the pre-training of the larger model. If reference and training models have different vocabularies, one can consider performing token alignment (Wan et al., 2024; Fu et al., 2023), which we leave for future work.
Weak-to-Strong Generation์€ ์ฐธ์กฐ ๋ฐ ์—ฐ์† ์‚ฌ์ „ ํ›ˆ๋ จ์— ๋™์ผํ•œ ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š” ์ฃผ์š” ์‹คํ—˜๊ณผ๋Š” ๋ณ„๋„๋กœ, ๋” ์ž‘์€ ์ฐธ์กฐ ๋ชจ๋ธ์ด ๋” ํฐ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ํšจ๊ณผ์ ์œผ๋กœ ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” Tinyllma-1.1B๋ฅผ ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  Llama-2-7B๋ฅผ ์ˆ˜ํ•™์— ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•œ๋‹ค. ํ‘œ 3์— ์ œ์‹œ๋œ ๊ฒฐ๊ณผ๋Š” ์†Œํ˜• ๋ชจ๋ธ๊ณผ ๋Œ€ํ˜• ๋ชจ๋ธ ์‚ฌ์ด์˜ ์ƒ๋‹นํ•œ ๊ฒฉ์ฐจ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ (Li ๋“ฑ, 2023), ํ† ํฐ ์„ ํƒ์— ์†Œํ˜• ์ฐธ์กฐ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํฐ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ›ˆ๋ จ์— ์—ฌ์ „ํžˆ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์ฐธ์กฐ ๋ชจ๋ธ๊ณผ ํ›ˆ๋ จ ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ์–ดํœ˜๋ฅผ ๊ฐ–๋Š” ๊ฒฝ์šฐ, ํ† ํฐ ์ •๋ ฌ(Wan ๋“ฑ, 2024; Fu ๋“ฑ, 2023)์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ–ฅํ›„ ์ž‘์—…์„ ์œ„ํ•ด ๋‚จ๊ฒจ๋‘”๋‹ค.
_You are a question-answer generator. Your goal is to generate question-answer pairs given the Context._
_๋‹น์‹ ์€ ์งˆ์˜์‘๋‹ต ์ƒ์„ฑ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋ชฉํ‘œ๋Š” ์ปจํ…์ŠคํŠธ๊ฐ€ ์ฃผ์–ด์ง„ ์งˆ๋ฌธ-์‘๋‹ต ์Œ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค._
**Paraphrase I** The mayoral election in Indianapolis took place on November 7, 2023, with preliminary elections occurring on May 2. The sitting mayor, Democrat Joe Hogsett, successfully ran for a third term. Both Hogsett and his Republican opponent, Jefferson Shreve, moved on to the main election.
**๋ฐ”๊พธ์–ด ๋งํ•˜๋ฉด I** ์ธ๋””์• ๋‚˜ํด๋ฆฌ์Šค์˜ ์‹œ์žฅ ์„ ๊ฑฐ๋Š” 2023๋…„ 11์›” 7์ผ์— ์‹ค์‹œ๋˜์—ˆ์œผ๋ฉฐ ์˜ˆ๋น„ ์„ ๊ฑฐ๋Š” 5์›” 2์ผ์— ์‹ค์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์ง ์‹œ์žฅ์ธ ๋ฏผ์ฃผ๋‹น ์กฐ ํ˜ธ๊ทธ์…‹์€ ์„ฑ๊ณต์ ์œผ๋กœ 3์„ ์— ์ถœ๋งˆํ–ˆ์Šต๋‹ˆ๋‹ค. ํ˜ธ๊ทธ์…‹๊ณผ ๊ทธ์˜ ๊ณตํ™”๋‹น ์ ์ˆ˜ ์ œํผ์Šจ ์Šˆ๋ฆฌ๋ธŒ๋Š” ๋ชจ๋‘ ๋ณธ์„ ๊ฑฐ๋กœ ๋„˜์–ด๊ฐ”๋‹ค.
When such generation process has been suspended by the scheduler at an intermediate step, the context manager uses the snapshot function to capture and store the current state of the LLM's beam search tree, including all intermediate probabilities and paths being explored for generating the response. Upon resumption, the restoration function is employed to reload the saved state from the snapshot, allowing the LLM to continue its generation process exactly from the point of suspension to reach the final answer: _Search weather in Paris_. In this way, the context manager ensures that the temporary suspension of one agent's request does not lead to a loss of progress, thereby optimizing resource use without compromising the quality and efficiency of response generation.
์ด๋Ÿฌํ•œ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์ค‘๊ฐ„ ๋‹จ๊ณ„์—์„œ ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ค‘๋‹จ๋˜์—ˆ์„ ๋•Œ, ์ปจํ…์ŠคํŠธ ๊ด€๋ฆฌ์ž๋Š” ์Šค๋ƒ…์ƒท ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ํƒ์ƒ‰๋˜๋Š” ๋ชจ๋“  ์ค‘๊ฐ„ ํ™•๋ฅ  ๋ฐ ๊ฒฝ๋กœ๋ฅผ ํฌํ•จํ•˜๋Š” LLM์˜ ๋น” ํƒ์ƒ‰ ํŠธ๋ฆฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์บก์ฒ˜ํ•˜๊ณ  ์ €์žฅํ•œ๋‹ค. ์žฌ์‹œ์ž‘ ์‹œ ๋ณต์› ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์Šค๋ƒ…์ƒท์—์„œ ์ €์žฅ๋œ ์ƒํƒœ๋ฅผ ๋‹ค์‹œ ๋กœ๋“œํ•˜์—ฌ LLM์ด ์ค‘์ง€ ์ง€์ ์—์„œ ์ •ํ™•ํžˆ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ณ„์†ํ•˜์—ฌ ์ตœ์ข… ๋‹ต๋ณ€์ธ _ํŒŒ๋ฆฌ์˜ ๋‚ ์”จ ๊ฒ€์ƒ‰_ ์— ๋„๋‹ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ, ์ปจํ…์ŠคํŠธ ๊ด€๋ฆฌ์ž๋Š” ํ•˜๋‚˜์˜ ์—์ด์ „ํŠธ์˜ ์š”์ฒญ์˜ ์ผ์‹œ์ ์ธ ์ค‘๋‹จ์ด ์ง„ํ–‰์˜ ์†์‹ค๋กœ ์ด์–ด์ง€์ง€ ์•Š๋„๋ก ๋ณด์žฅํ•จ์œผ๋กœ์จ, ์‘๋‹ต ์ƒ์„ฑ์˜ ํ’ˆ์งˆ ๋ฐ ํšจ์œจ์„ฑ์„ ์†์ƒ์‹œํ‚ค์ง€ ์•Š์œผ๋ฉด์„œ ์ž์› ์‚ฌ์šฉ์„ ์ตœ์ ํ™”ํ•œ๋‹ค.
Table 2 provides insights into the characteristics of the three datasets involved in our experimental procedure, including the total number of questions and the average number of words per ques
ํ‘œ 2๋Š” ์ด ์งˆ๋ฌธ ์ˆ˜์™€ ์งˆ๋ฌธ๋‹น ํ‰๊ท  ๋‹จ์–ด ์ˆ˜๋ฅผ ํฌํ•จํ•˜์—ฌ ์‹คํ—˜ ์ ˆ์ฐจ์— ๊ด€๋ จ๋œ ์„ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํŠน์„ฑ์— ๋Œ€ํ•œ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•œ๋‹ค.
Adopting open source models as reference models could be dangerous.We also show mean \(\gamma_{\pi_{ref}}\) of \(M_{2}\), the aligned model from step 2 of sDPO. Unlike \(S\) and \(M_{1}\), \(M_{2}\) is trained on the Ultrafeedback Cleaned dataset, _i.e.,_\(M_{2}\) is used as a reference model on data that was already used to train it. Note that such a case could happen commonly when adopting various open source models as reference models. This is because the datasets that were used in training those models are often unclear and could overlap with the preference datasets unintentionally. Mean \(\gamma_{\pi_{ref}}\) of \(M_{2}\) is \(84.35\), which is staggeringly higher than either \(S\) or \(M_{1}\). The strikingly high value for \(M_{2}\) likely points to overfitting of \(M_{2}\) to the Ultrafeedback Cleaned dataset. This result highlights the potential danger of merely adopting open source models as reference models instead of using sDPO.
์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•˜๋Š” ๊ฒƒ์€ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋˜ํ•œ sDPO์˜ 2๋‹จ๊ณ„๋ถ€ํ„ฐ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ธ \(M_{2}\)์˜ ํ‰๊ท  \(\gamma_{\pi_{ref}}\)์„ ๋ณด์—ฌ์ค€๋‹ค. \(S\) ๋ฐ \(M_{1}\)์™€ ๋‹ฌ๋ฆฌ \(M_{2}\)๋Š” Ultrafeedback Cleaned ๋ฐ์ดํ„ฐ์…‹์—์„œ ํ•™์Šต๋˜๋ฉฐ _์ฆ‰,_\(M_{2}\)๋Š” ์ด๋ฏธ ํ•™์Šต์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋œ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ๋Š” ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•  ๋•Œ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ข…์ข… ๋ถˆ๋ถ„๋ช…ํ•˜๊ณ  ์˜๋„ํ•˜์ง€ ์•Š๊ฒŒ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฒน์น  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. \(M_{2}\)์˜ ํ‰๊ท  \(\gamma_{\pi_{ref}}\)๋Š” \(84.35\)๋กœ \(S\) ๋˜๋Š” \(M_{1}\)๋ณด๋‹ค ๋งค์šฐ ๋†’๋‹ค. \(M_{2}\)์— ๋Œ€ํ•œ ํ˜„์ €ํ•˜๊ฒŒ ๋†’์€ ๊ฐ’์€ \(M_{2}\)๊ฐ€ ์šธํŠธ๋ผํ”ผ๋“œ๋ฐฑ ์ฒญ์†Œ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ณผ์ ํ•ฉ๋˜์—ˆ์Œ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” sDPO๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•˜๋Š” ๊ฒƒ์˜ ์ž ์žฌ์  ์œ„ํ—˜์„ ๊ฐ•์กฐํ•œ๋‹ค.
- Participating in the debate would give Trump a platform to outline his policies and vision, which could be beneficial for his campaign.
ํ† ๋ก ํšŒ์— ์ฐธ์—ฌํ•˜๋ฉด ํŠธ๋Ÿผํ”„์—๊ฒŒ ์ž์‹ ์˜ ์ •์ฑ…๊ณผ ๋น„์ „์„ ๊ฐœ๊ด„ํ•  ์ˆ˜ ์žˆ๋Š” ํ”Œ๋žซํผ์ด ๋งˆ๋ จ๋ผ ์„ ๊ฑฐ ์šด๋™์— ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค.
Figure A20: Comparison of data selection strategies with the OPT model embedding space, when using D4 as a the selection strategy, when using C4 as the starting training dataset. The x-axis is selectoin ratio \(R\), and the y-axis is perplexity difference compared to baseline (the horizontal gray dotted line at 0.0 represents our baseline i.e. when no data selection is done), so **lower is better**. Notice that D4 and SemDeDup match at 90%, because we use \(R_{dedup}=0.9\) and varied \(R_{proto}\) for this experiment.
๊ทธ๋ฆผ A20: ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต๊ณผ OPT ๋ชจ๋ธ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์˜ ๋น„๊ต, ์„ ํƒ ์ „๋žต์œผ๋กœ D4๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ, ์‹œ์ž‘ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ C4๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ. x์ถ•์€ ์„ ํƒ ๋น„์œจ \(R\)์ด๊ณ  y์ถ•์€ ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋ณต์žก๋„ ์ฐจ์ด(0.0์˜ ์ˆ˜ํ‰ ํšŒ์ƒ‰ ์ ์„ ์€ ๊ธฐ์ค€์„ , ์ฆ‰ ๋ฐ์ดํ„ฐ ์„ ํƒ์ด ์ˆ˜ํ–‰๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ)์ด๋ฏ€๋กœ **๋‚ฎ์Œ**์ด ๋” ์ข‹์Šต๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์„ ์œ„ํ•ด \(R_{dedup}=0.9\)์„ ์‚ฌ์šฉํ•˜๊ณ  \(R_{proto}\)์„ ๋ณ€ํ™”์‹œ์ผฐ๊ธฐ ๋•Œ๋ฌธ์— D4์™€ SemDeDup์ด 90% ์ผ์น˜ํ•จ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.
**RR** proposed using autoregressive reward models in discussions with **EM**; derived the DPO objective; proved the theoretical properties of the algorithm and wrote the relevant sections and appendices. He also suggested and helped with organizing experiments and contributed some of the PPO and reward learning baselines.
**EM**๊ณผ์˜ ๋…ผ์˜์—์„œ ์ž๊ธฐํšŒ๊ท€ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ œ์•ˆ๋œ **RR**, DPO ๋ชฉํ‘œ๋ฅผ ๋„์ถœ, ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ด๋ก ์  ์†์„ฑ์„ ์ฆ๋ช…ํ•˜๊ณ  ๊ด€๋ จ ์„น์…˜ ๋ฐ ๋ถ€๋ก์„ ์ž‘์„ฑํ–ˆ๋‹ค. ๊ทธ๋Š” ๋˜ํ•œ ์‹คํ—˜์„ ์กฐ์งํ•˜๋Š” ๊ฒƒ์„ ์ œ์•ˆํ•˜๊ณ  ๋„์™”๊ณ  PPO์™€ ๋ณด์ƒ ํ•™์Šต ๊ธฐ์ค€์„ ์˜ ์ผ๋ถ€๋ฅผ ๊ธฐ์—ฌํ–ˆ๋‹ค.
We conduct experiments on different model sizes. First, to show the effectiveness of stage-2 context-enhanced instruction tuning, we compare against the **Llama2-SFT-7B/13B/70B**, which is the Llama2-7B/13B/70B foundation model after the stage-1 supervised fine-tuning (SFT). Second, we compare against **Llama2-Chat-7B/13B/70B** since Llama2-Chat models are shown to possess strong instruction following and conversational QA capabilities (Touvron et al., 2023). Aside from Llama2 models, we also conduct experiments on our in-house GPT-8B foundation model, and we compare against its stage-1 SFT baseline (**GPT-8B-SFT**). Finally, we compare against two very strong OpenAI models: **GPT-3.5-turbo (4k)** and **GPT-4 (8k)**. For fair comparison, we use the same context as inputs for both our models and baselines. Note that we have carefully tuned the instructions for all the baselines to ensure they achieve as good as possible results.3
์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•œ๋‹ค. ๋จผ์ € ๋‹จ๊ณ„ 2 ์ปจํ…์ŠคํŠธ ๊ฐ•ํ™” ๋ช…๋ น ํŠœ๋‹์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด ๋‹จ๊ณ„ 1 ๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ • (SFT) ํ›„ Llama2-7B/13B/70B ๊ธฐ์ดˆ ๋ชจ๋ธ์ธ **Llama2-SFT-7B/13B/70B** ์™€ ๋น„๊ต ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, Llama2-Chat ๋ชจ๋ธ์€ ๋‹ค์Œ ๋ฐ ๋Œ€ํ™” QA ๊ธฐ๋Šฅ์ด ๊ฐ•๋ ฅํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๊ธฐ ๋•Œ๋ฌธ์— **Llama2-Chat-7B/13B/70B** ์™€ ๋น„๊ต ํ•ฉ๋‹ˆ๋‹ค (Touvron ๋“ฑ, 2023). Llama2 ๋ชจ๋ธ ์™ธ์—๋„ ์‚ฌ๋‚ด GPT-8B ๊ธฐ์ดˆ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹คํ—˜๋„ ์ˆ˜ํ–‰ํ•˜๊ณ  1๋‹จ๊ณ„ SFT ๊ธฐ์ค€์„ (**GPT-8B-SFT**)๊ณผ ๋น„๊ตํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๋‘ ๊ฐœ์˜ ๋งค์šฐ ๊ฐ•๋ ฅํ•œ OpenAI ๋ชจ๋ธ์ธ **GPT-3.5-ํ„ฐ๋ณด(4k)** ๋ฐ **GPT-4(8k)** ์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ๊ณผ ๊ธฐ์ค€์„  ๋ชจ๋‘์— ๋Œ€ํ•œ ์ž…๋ ฅ๊ณผ ๋™์ผํ•œ ์ปจํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ๋ชจ๋“  ๊ธฐ์ค€์„ ์ด ๊ฐ€๋Šฅํ•œ ํ•œ ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋„๋ก ์ฃผ์˜ ๊นŠ๊ฒŒ ์ง€์นจ์„ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. 3
Figure 6: **Compute-Normalized Generalization Results**. We visualize the performance curves normalized by the number of inference calls used.
๊ทธ๋ฆผ 6: **Compute-Normalized Generalization Results**. ์‚ฌ์šฉ๋œ ์ถ”๋ก  ํ˜ธ์ถœ์˜ ์ˆ˜๋กœ ์ •๊ทœํ™”๋œ ์„ฑ๋Šฅ ๊ณก์„ ์„ ์‹œ๊ฐํ™”ํ•œ๋‹ค.
After the release of ChatGPT (OpenAI, 2022), instruction tuning becomes an indispensable ingredient to build the state-of-the-art dialogue agent which has astonishing zero shot capability to a wide range of tasks. The conversational QA capability plays a crucial role in a dialogue agent, yet there has been limited research dedicated to this vital aspect.
ChatGPT(OpenAI, 2022)๊ฐ€ ์ถœ์‹œ๋œ ํ›„, ๋ช…๋ น์–ด ํŠœ๋‹์€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์— ๋†€๋ผ์šด ์ œ๋กœ ์ƒท ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ์ตœ์ฒจ๋‹จ ๋Œ€ํ™” ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ํ•„์ˆ˜ ์š”์†Œ๊ฐ€ ๋œ๋‹ค. ๋Œ€ํ™”ํ˜• QA ๋Šฅ๋ ฅ์€ ๋Œ€ํ™” ์—์ด์ „ํŠธ์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ์ค‘์š”ํ•œ ์ธก๋ฉด์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฐ๊ตฌ๋Š” ์ œํ•œ์ ์ด๋‹ค.
After SFT, we employ DPO (Rafailov et al., 2024) for human preference alignment of the model. During this stage, UltraFeedback (Cui et al., 2023) is utilized as the primary alignment dataset, and an proprietary preference dataset is constructed to enhance the model's code and mathematical capabilities. We conduct one epoch of DPO training with a learning rate of \(1\times 10^{-5}\) and utilize a Cosine LRS since we have a pre-defined training steps.
SFT ์ดํ›„ ๋ชจ๋ธ์˜ ์ธ๊ฐ„ ์„ ํ˜ธ๋„ ์ •๋ ฌ์„ ์œ„ํ•ด DPO(Rafailov et al., 2024)๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์ด ๋‹จ๊ณ„ ๋™์•ˆ, UltraFeedback(Cui et al., 2023)์ด ์ผ์ฐจ ์ •๋ ฌ ๋ฐ์ดํ„ฐ์„ธํŠธ๋กœ ํ™œ์šฉ๋˜๋ฉฐ, ๋ชจ๋ธ์˜ ์ฝ”๋“œ ๋ฐ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋…์  ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์„ธํŠธ๊ฐ€ ๊ตฌ์ถ•๋œ๋‹ค. ํ•™์Šต๋ฅ ์ด \(1\times 10^{-5}\)์ธ DPO ํ›ˆ๋ จ์˜ ํ•œ ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›ˆ๋ จ ๋‹จ๊ณ„๊ฐ€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ฝ”์‚ฌ์ธ LRS๋ฅผ ํ™œ์šฉํ•œ๋‹ค.
We compare our method against state-of-the-art data generation approaches for instruction tuning. For fair comparison, we provide all methods the same LLM backbones when possible. Moreover, we control the number of instruction-response pairs the same for all methods to ablate the effect of data quantity. Baseline methods include **Self-Instruct**Wang et al. (2022), **Alpagasus**Chen et al. (2023), **Tree-Instruct**, **WizardLM**Xu et al. (2023), and **WizardLM+**, an enhanced version of WizardLM using the same basic instructions generated from CodecLM as seed instructions. Baseline details are presented in Appendix A.2.
๋ณธ ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ช…๋ น์–ด ํŠœ๋‹์„ ์œ„ํ•œ ์ตœ์‹  ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ•œ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ํ•œ ๋ชจ๋“  ๋ฐฉ๋ฒ•์— ๋™์ผํ•œ LLM ๋ฐฑ๋ณธ์„ ์ œ๊ณตํ•œ๋‹ค. ๋˜ํ•œ, ๋ฐ์ดํ„ฐ ์–‘์˜ ์˜ํ–ฅ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋“  ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๋ช…๋ น-์‘๋‹ต ์Œ์˜ ์ˆ˜๋ฅผ ๋™์ผํ•˜๊ฒŒ ์ œ์–ดํ•œ๋‹ค. ๊ธฐ๋ณธ ๋ฐฉ๋ฒ•์—๋Š” CodecLM์—์„œ ์ƒ์„ฑ ๋œ ๊ธฐ๋ณธ ์ง€์นจ์„ ์‹œ๋“œ ์ง€์นจ์œผ๋กœ ์‚ฌ์šฉ ํ•˜ ์—ฌ ๊ฐ•ํ™” ๋œ ๋ฒ„์ „์˜ WizardLM์ธ **Self-Instruct**Wang ๋“ฑ (2022), **Alpagasus**Chen ๋“ฑ (2023), **Tree-Instruct**, **WizardLM**Xu ๋“ฑ (2023) ๋ฐ **WizardLM+** ๊ฐ€ ํฌํ•จ ๋ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€ ์„ธ๋ถ€ ์ •๋ณด๋Š” ๋ถ€๋ก A.2์— ๋‚˜์™€ ์žˆ๋‹ค.
For the 100-exposure scaling law, we additionally compare three configurations: \((topk,cap\_factor)=(1,2),(2,1),(2,2)\), finding minimal differences among them as shown in Figure 17. Remember from Section 7 that differences in model architecture usually become apparent in the insufficient training regime; this is why we opt for 100-exposure instead of 1000-exposure. Notably, \((topk,cap\_factor)=(2,2)\) performs best (among the three) for deep models, such as GPT2-16-4 with 32 experts.
100-๋…ธ์ถœ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ฒฝ์šฐ, ๊ทธ๋ฆผ 17๊ณผ ๊ฐ™์ด \((topk,cap\_factor)=(1,2),(2,1),(2,2)\)์˜ ์„ธ ๊ฐ€์ง€ ๊ตฌ์„ฑ์„ ์ถ”๊ฐ€๋กœ ๋น„๊ตํ•˜๋ฉฐ, ์ด๋“ค ์‚ฌ์ด์˜ ์ตœ์†Œ ์ฐจ์ด๋ฅผ ์ฐพ๋Š”๋‹ค. ์„น์…˜ 7์—์„œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์˜ ์ฐจ์ด๋Š” ๋ณดํ†ต ๋ถˆ์ถฉ๋ถ„ํ•œ ํ›ˆ๋ จ ์ฒด์ œ์—์„œ ๋ช…๋ฐฑํ•ด์ง„๋‹ค๋Š” ๊ฒƒ์„ ๊ธฐ์–ตํ•˜๋ผ; ์ด๊ฒƒ์ด ์šฐ๋ฆฌ๊ฐ€ 1000-๋…ธ์ถœ ๋Œ€์‹ ์— 100-๋…ธ์ถœ์„ ์„ ํƒํ•˜๋Š” ์ด์œ ์ด๋‹ค. ํŠนํžˆ \((topk,cap\_factor)=(2,2)\)๋Š” 32๋ช…์˜ ์ „๋ฌธ๊ฐ€์™€ ํ•จ๊ป˜ GPT2-16-4์™€ ๊ฐ™์€ ์‹ฌ์ธต ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค.
To validate our hypothesis, we conducted several experiments on the PopQA dataset [22], as it provides a suitable platform for benchmarking hybrid retrieval strategies. As a result of these experiments we find that:
๊ฐ€์„ค์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ๊ฒ€์ƒ‰ ์ „๋žต์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•œ ํ”Œ๋žซํผ์„ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์— PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ [22]์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹คํ—˜๋“ค์˜ ๊ฒฐ๊ณผ๋กœ์„œ ์šฐ๋ฆฌ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•œ๋‹ค:
\begin{tabular}{p{56.9pt}|p{399.8pt}} \hline **ChatGPT** & Training a LLM (Language Model with Learning) using RLHF (Reinforcement Learning with Human Feedback) methods based on InstructGPT can be a complex process that requires expertise in both natural language processing and machine learning. Here are some general steps you can take to create a human assistant that exceeds open assistants in performance, accuracy, and utility:
\begin{tabular}{p{56.9pt}|p{399.8pt}} \hline InstructGPT ๊ธฐ๋ฐ˜ RLHF (Reinforcement Learning with Human Feedback) ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉ ํ•˜ ์—ฌ LLM (Language Model with Learning)์„ ํ•™์Šต ํ•˜๋Š” ๊ฒƒ์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ฐ ๊ธฐ๊ณ„ ํ•™์Šต ๋ชจ๋‘์— ์ „๋ฌธ ์ง€์‹์ด ํ•„์š”ํ•œ ๋ณต์žกํ•œ ํ”„๋กœ์„ธ์Šค์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์„ฑ๋Šฅ, ์ •ํ™•๋„ ๋ฐ ์œ ํ‹ธ๋ฆฌํ‹ฐ์—์„œ ๊ฐœ๋ฐฉํ˜• ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์ดˆ๊ณผํ•˜๋Š” ์ธ๊ฐ„ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ทจํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ๋‹จ๊ณ„๊ฐ€ ์žˆ๋‹ค:
The story often matches back to Jurning childhood, producing context about her children and the relationship with my.
๊ทธ ์ด์•ผ๊ธฐ๋Š” ์ข…์ข… ์ €๋‹์˜ ์–ด๋ฆฐ ์‹œ์ ˆ๊ณผ ์ผ์น˜ํ•˜์—ฌ ๊ทธ๋…€์˜ ์•„์ด๋“ค๊ณผ ๋‚˜์˜ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ์ƒ์‚ฐํ•œ๋‹ค.
- Given the recent escalation in tensions and historical precedents, there is a non-negligible chance that Israel could conduct a military operation against Iran.
์ตœ๊ทผ ๊ธด์žฅ์˜ ๊ณ ์กฐ์™€ ์—ญ์‚ฌ์  ์ „๋ก€๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์ด์Šค๋ผ์—˜์ด ์ด๋ž€์— ๋Œ€ํ•œ ๊ตฐ์‚ฌ ์ž‘์ „์„ ์ˆ˜ํ–‰ํ•  ๊ฐ€๋Šฅ์„ฑ์€ ๋ฌด์‹œํ•  ์ˆ˜ ์—†๋‹ค.
MAMmnoTH Yue et al. (2023) introduces the data named MathInstruct incorporates a variety of math problem datasets, including but not limited to GSM8K, MATH, AQUA Ling et al. (2017), and Camel-Math Li et al. (2023). It also leverages GPT-4 to generate new reasoning paths through both Program-of-Thought (POT) Chen et al. (2022) and Chain-of-Thought (COT) methods.
MAMmnoTH Yue ๋“ฑ(2023)์€ GSM8K, MATH, AQUA Ling ๋“ฑ(2017) ๋ฐ Camel-Math Li ๋“ฑ(2023)์„ ํฌํ•จํ•˜์ง€๋งŒ ์ด์— ์ œํ•œ๋˜์ง€ ์•Š๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ ๋ฌธ์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” MathInstruct๋ผ๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์†Œ๊ฐœํ•œ๋‹ค. ๋˜ํ•œ GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ Chen et al.(2022)๊ณผ Chain-of-Think(COT) ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ์„ฑํ•œ๋‹ค.
While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper we introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss. The resulting algorithm, which we call _Direct Preference Optimization_ (DPO), is stable, performant, and computationally lightweight, eliminating the need for sampling from the LM during fine-tuning or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.
๋Œ€๊ทœ๋ชจ ๋น„๊ฐ๋… ์–ธ์–ด ๋ชจ๋ธ(LMs)์€ ๊ด‘๋ฒ”์œ„ํ•œ ์„ธ๊ณ„ ์ง€์‹๊ณผ ์ผ๋ถ€ ์ถ”๋ก  ๊ธฐ์ˆ ์„ ๋ฐฐ์šฐ๋Š” ๋ฐ˜๋ฉด, ํ›ˆ๋ จ์˜ ์™„์ „ํ•œ ๋น„๊ฐ๋… ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ํ–‰๋™์˜ ์ •ํ™•ํ•œ ์ œ์–ด๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๋‹ค. ์ด๋Ÿฌํ•œ ์กฐํ–ฅ์„ฑ์„ ์–ป๊ธฐ ์œ„ํ•œ ๊ธฐ์กด์˜ ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ ์„ธ๋Œ€์˜ ์ƒ๋Œ€์ ์ธ ํ’ˆ์งˆ์˜ ์ธ๊ฐ„ ๋ผ๋ฒจ์„ ์ˆ˜์ง‘ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๊ฐ๋…๋˜์ง€ ์•Š์€ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋ฉฐ, ์ข…์ข… ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ(RLHF)์œผ๋กœ๋ถ€ํ„ฐ์˜ ๊ฐ•ํ™” ํ•™์Šต๊ณผ ํ•จ๊ป˜ ์กฐ์ •ํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RLHF๋Š” ๋ณต์žกํ•˜๊ณ  ์ข…์ข… ๋ถˆ์•ˆ์ •ํ•œ ์ ˆ์ฐจ๋กœ, ๋จผ์ € ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ๋ณด์ƒ ๋ชจ๋ธ์„ ํ”ผํŒ…ํ•œ ๋‹ค์Œ ๊ฐ•ํ™” ํ•™์Šต์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์—์„œ ๋„ˆ๋ฌด ๋ฉ€๋ฆฌ ํ‘œ๋ฅ˜ํ•˜์ง€ ์•Š๊ณ  ์ถ”์ •๋œ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฐ„๋‹จํ•œ ๋ถ„๋ฅ˜ ์†์‹ค๋งŒ์œผ๋กœ ํ‘œ์ค€ RLHF ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ํ์‡„ํ˜• ์ตœ์  ์ •์ฑ… ์ถ”์ถœ์ด ๊ฐ€๋Šฅํ•œ RLHF ๋ณด์ƒ ๋ชจ๋ธ์˜ ์ƒˆ๋กœ์šด ํŒŒ๋ผ๋ฏธํ„ฐํ™”๋ฅผ ์†Œ๊ฐœํ•œ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ _์ง์ ‘ ์„ ํ˜ธ ์ตœ์ ํ™”_(DPO)๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ๊ฒฐ๊ณผ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์•ˆ์ •์ ์ด๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ ๊ณ„์‚ฐ์ ์œผ๋กœ ๊ฐ€๋ฒผ์›Œ ๋ฏธ์„ธ ์กฐ์ • ๋˜๋Š” ์ƒ๋‹นํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋™์•ˆ LM์—์„œ ์ƒ˜ํ”Œ๋งํ•  ํ•„์š”๊ฐ€ ์—†๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ DPO๊ฐ€ ๊ธฐ์กด์˜ ๋ฐฉ๋ฒ•๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์™€ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ๋” ๋‚˜์€ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค. ํŠนํžˆ, DPO์™€์˜ ๋ฏธ์„ธ ์กฐ์ •์€ ์„ธ๋Œ€์˜ ๊ฐ์„ฑ์„ ์ œ์–ดํ•˜๋Š” ๋Šฅ๋ ฅ์—์„œ PPO ๊ธฐ๋ฐ˜ RLHF๋ฅผ ์ดˆ๊ณผํ•˜๊ณ , ๊ตฌํ˜„ ๋ฐ ํ›ˆ๋ จ์ด ์ƒ๋‹นํžˆ ๊ฐ„๋‹จํ•˜๋ฉด์„œ ์š”์•ฝ ๋ฐ ์‹ฑ๊ธ€ ํ„ด ๋Œ€ํ™”์—์„œ ์‘๋‹ต ํ’ˆ์งˆ์„ ์ผ์น˜์‹œํ‚ค๊ฑฐ๋‚˜ ๊ฐœ์„ ํ•œ๋‹ค.
In comparing the performance of Nemotron-4 15B on TyDiQA-GoldP to a range of models, Table 8 shows that Nemotron-4 15B achieves the best performance. Impressively, Nemotron-4 15B is able to significantly improve upon the next best model, PaLM 62B-cont.
TyDiQA-GoldP ์ƒ์˜ ๋„ค๋ชจํŠธ๋ก -4 15B์˜ ์„ฑ๋Šฅ์„ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•  ๋•Œ, ํ‘œ 8์€ ๋„ค๋ชจํŠธ๋ก -4 15B๊ฐ€ ์ตœ์ƒ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•จ์„ ๋ณด์—ฌ์ค€๋‹ค. ์ธ์ƒ์ ์œผ๋กœ, ๋„ค๋ชจํŠธ๋ก -4 15B๋Š” ์ฐจ์„ ์ฑ…์ธ PaLM 62B-cont์—์„œ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค.
RealNewsis a subset of the Common Crawl consisting of articles from news domains Zellers et al. (2019). It contains 31M documents with average length 793 BPE tokens. RealNews was deduplicated by inserting a hash of the first 100 characters of each document into a bloom filter Bloom (1970) and then excluding any document which resulted in a hash collision. Like C4, examples with duplicate URLs were excluded.
RealNewsis a subset of the Common Crawl consisting to articles from news domains Zellers et al. (2019). ํ‰๊ท  ๊ธธ์ด 793 BPE ํ† ํฐ์„ ๊ฐ€์ง„ 31M ๋ฌธ์„œ๊ฐ€ ๋“ค์–ด ์žˆ์Šต๋‹ˆ๋‹ค. RealNews๋Š” Bloom ํ•„ํ„ฐ Bloom(1970)์— ๊ฐ ๋ฌธ์„œ์˜ ์ฒ˜์Œ 100์ž์˜ ํ•ด์‹œ๋ฅผ ์‚ฝ์ž…ํ•œ ํ›„ ํ•ด์‹œ ์ถฉ๋Œ์ด ๋ฐœ์ƒํ•œ ๋ฌธ์„œ๋ฅผ ์ œ์™ธํ•จ์œผ๋กœ์จ ์ค‘๋ณต ์ œ๊ฑฐ๋˜์—ˆ๋‹ค. C4์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ค‘๋ณต URL์ด ์žˆ๋Š” ์˜ˆ๋Š” ์ œ์™ธ๋˜์—ˆ๋‹ค.
* A model trained from random initialization on the union of all datasets i.e. \(\bigcup_{i=0}^{N-1}\mathcal{D}_{i}\), and * A model trained from random initialization on individual dataset \(\mathcal{D}_{i}\), \(0\leq i\leq N\).
* ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐํ•ฉ์— ๋Œ€ํ•œ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋กœ๋ถ€ํ„ฐ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ, ์ฆ‰ \(\bigcup_{i=0}^{N-1}\mathcal{D}_{i}\), ๋ฐ * ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{i}\), \(0\leq i\leq N\)์— ๋Œ€ํ•œ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.
1. To make the parsing of each step easier, we few-shot re-generate 15k answers for GSM8k and MATH with an Alpha version of WizardLM 70B model to produce solutions in a step-by-step format, then find out those with a correct answer, and use this data to finetune base Llama model. 2. To enhance the model's ability to adhere to the neural and diverse instructions, we also sample 1.5k open-domain conversations from WizardLM's training data, then merge it with above math corpus as the final SFT training data.
1. ๊ฐ ๋‹จ๊ณ„์˜ ๊ตฌ๋ฌธ ๋ถ„์„์„ ๋” ์‰ฝ๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด GSM8k ๋ฐ MATH์— ๋Œ€ํ•ด Alpha ๋ฒ„์ „์˜ WizardLM 70B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ 15k ๋‹ต๋ณ€์„ ๊ฑฐ์˜ ์ƒท ์žฌ์ƒ์„ฑํ•˜์—ฌ ๋‹จ๊ณ„๋ณ„ ํ˜•์‹์œผ๋กœ ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์ •๋‹ต์„ ์ฐพ์€ ๋‹ค์Œ ์ด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ผ๋งˆ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•œ๋‹ค. 2. ๋ชจ๋ธ์˜ ์‹ ๊ฒฝ ๋ฐ ๋‹ค์–‘ํ•œ ์ง€์นจ์„ ์ค€์ˆ˜ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์œ„์ €๋“œ๋กฌ์˜ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ 1.5k ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋Œ€ํ™”๋ฅผ ์ƒ˜ํ”Œ๋งํ•œ ๋‹ค์Œ ์ตœ์ข… SFT ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋กœ ์œ„์˜ ์ˆ˜ํ•™ ๋ง๋ญ‰์น˜์™€ ๋ณ‘ํ•ฉํ•œ๋‹ค.