YC-Chen's picture
Update README.md
ec61a06 verified
|
raw
history blame
16.7 kB
metadata
pipeline_tag: text-generation
license: apache-2.0
language:
  - zh

Model Card for Breeze-7B-Instruct-v0.1

Breeze-7B is a language model that builds upon the foundation of Mistral-7B-v0.1, specifically enhanced for Traditional Chinese.

Breeze-7B-Base-v0.1 introduces an expanded vocabulary with additional 30,000 Traditional Chinese tokens and is pre-trained on a substantial dataset of 250GB of Traditional Chinese content. With the expanded vocabulary, the base model operates at twice the inference speed for Traditional Chinese characters compared to Mistral-7B. [See Inference Performance.] This achievement marks a significant milestone as it is the first instance of vocabulary expansion in a model tailored for Traditional Chinese.

Breeze-7B-Instruct-v0.1 derives from the base model Breeze-7B-Base-v0.1 and has undergone supervised fine-tuning with over 1 million instances to sharpen its capabilities. This fine-tuned model demonstrates impressive performance in benchmarks for both English and Traditional Chinese, surpassing the results of Taiwan-LLM-7B-v2.1-chat, Taiwan-LLM-13B-v2.0-chat and Qwen-7B-chat in Traditional Chinese assessments. It also excels in some benchmarks against Yi-6B-Chat. In English evaluations, Breeze-7B-Instruct-v0.1 shows comparable results to Mistral-7B-Instruct-v0.1 on the MMLU and MT-Bench benchmarks. [See Chat Model Performance.]

Breeze-7B-Instruct-64k-v0.1 is an extension to Breeze-7B-Instruct-v0.1 to enable 64k context length, which is equivalent to 88k Traditional Chinese characters. With minimal sacrifice in the performance of the regular benchmarks, Breeze-7B-Instruct-64k-v0.1 can solve tasks such as question answering and summarization on document-level inputs. [See Long-context Performance.]

A project by the members (in alphabetical order): Chan-Jan Hsu 許湛然, Chang-Le Liu 劉昶樂, Feng-Ting Liao 廖峰挺, Po-Chun Hsu 許博竣, Yi-Chang Chen 陳宜昌, and the supervisor Da-Shan Shiu 許大山.

Features

  • Breeze-7B-Base-v0.1
    • Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
    • 8k tokens context length
  • Breeze-7B-Instruct-v0.1
    • Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
    • 8k tokens context length
    • Multi-turn dialogue (without special handling for harmfulness)
  • Breeze-7B-Instruct-64k-v0.1
    • Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
    • 64k tokens context length
    • Multi-turn dialogue (without special handling for harmfulness)

Model Details

  • Breeze-7B-Base-v0.1
    • Finetuned from: mistralai/Mistral-7B-v0.1
    • Model type: Causal decoder-only transformer language model
    • Language: English and Traditional Chinese (zh-tw)
  • Breeze-7B-Instruct-v0.1
  • Breeze-7B-Instruct-64k-v0.1

Base Model Performance

TMMLU+, DRCD, and Table source from MediaTek-Research/TCEval-v2. MediaTek-Research/TCEval-v2 derives from TCEval-v1 and ikala/tmmluplus. MMLU sources from hails/mmlu_no_train. We use the code revised from EleutherAI/lm-evaluation-harness to evaluate TMMLU+, DRCD, Table, and MMLU.

Models ↑ TMMLU+ (ACC) DRCD (EM) Table (ACC) MMLU (ACC)
TC, Knowledge TC, Reasoning TC, Reasoning EN, Knowledge
5 shot 3 shot 5 shot 5 shot
Yi-34B 34B 63.10 84.57 49.31 77.42
Qwen-14B 14B 51.30 16.95 * 50.69 68.83
Yi-6B 6B 49.63 76.61 34.72 65.35
Qwen-7B 7B 42.84 0.0 * 39.58 61.00
Breeze-7B-Base-v0.1 7B 40.35 81.13 28.47 61.63
Mistral-7B-v0.1 7B 36.93 79.27 27.78 64.89

* Few-shot learning cannot effectively guide the model to generate the proper answer.

Category ACC of TMMLU+ (5 shot)

Models STEM Social Science Humanities Other ↑ AVG
Yi-34B 56.03 73.06 61.12 62.19 63.10
Qwen-14B 46.51 58.20 51.12 49.38 51.30
Yi-6B 41.14 57.77 50.22 49.39 49.63
Qwen-7B 28.25 47.80 43.14 42.17 42.84
Breeze-7B-Base-v0.1 35.74 46.08 40.29 39.27 40.35
Mistral-7B-v0.1 33.01 42.23 35.86 37.63 36.93

Chat Model Performance

TMMLU+, DRCD, Table, and MT-Bench-tw source from MediaTek-Research/TCEval-v2. MediaTek-Research/TCEval-v2 derives from TCEval-v1 and ikala/tmmluplus. MMLU sources from hails/mmlu_no_train. MT-Bench source from lmsys/mt_bench_human_judgments. We use the code revised from EleutherAI/lm-evaluation-harness to evaluate TMMLU+, DRCD, Table, and MMLU. We use the code revised from fastchat llm_judge (GPT4 as judge) to evaluate MT-Bench-tw and MT-Bench.

Models ↑ MT-Bench-tw (Score) TMMLU+ (ACC) TMMLU+ (ACC) DRCD (EM) Table (ACC) MT-Bench (Score) MMLU (ACC) MMLU (ACC)
TC, Chat TC, Knowledge TC, Knowledge TC, Reasoning TC, Reasoning EN, Chat EN, Knowledge EN, Knowledge
0 shot 0 shot 5 shot 3 shot 0 shot 0 shot 0 shot 5 shot
gpt-3.5-turbo 7.1 41.76 7.9 70.00
Yi-34B-Chat 34B 6.9 54.87 36.81 7.6 71.04
Qwen-14B-Chat 14B 6.4 48.41 41.67 7.2 64.91
Breeze-7B-Instruct-v0.1 7B 5.7 41.61 45.83 7.1 63.26
Breeze-7B-Instruct-64k-v0.1 7B 5.5 40.99 36.11 7.1 63.68
Qwen-7B-Chat 7B 5.4 40.02 33.33 6.2 55.94
Yi-6B-Chat 6B 5.0 44.79 25.69 6.0 59.45
Taiwan-LLM-13B-v2.0-chat 13B 5.0 29.47 23.61 -* 50.50
Taiwan-LLM-7B-v2.1-chat 7B 4.2 28.08 31.25 -* 42.72

* Taiwan-LLM models responds to multi-turn questions (English) in Traditional Chinese.

Category Score of MT-Bench-tw (0 shot)

Models STEM Extraction Reasoning Math Coding Roleplay Writing Humanities ↑ AVG
gpt-3.5-turbo
Yi-34B-Chat
Qwen-14B-Chat
Breeze-7B-Instruct-v0.1
Breeze-7B-Instruct-64k-v0.1
Qwen-7B-Chat
Yi-6B-Chat
Taiwan-LLM-13B-v2.0-chat
Taiwan-LLM-7B-v2.1-chat

Category ACC of TMMLU+ (0 shot)

Model STEM Social Science Humanities Other ↑ AVG
Yi-34B-Chat 47.65 64.25 52.73 54.91 54.87
Qwen-14B-Chat 43.83 55.00 48.55 46.22 48.41
Yi-6B-Chat 37.80 51.74 45.36 44.25 44.79
gpt-3.5-turbo 41.56 46.72 36.73 42.03 41.76
Breeze-7B-Instruct-v0.1 37.41 46.81 42.06 40.16 41.61
Breeze-7B-Instruct-64k-v0.1 37.88 46.35 40.31 39.40 40.99
Qwen-7B-Chat 35.44 46.22 38.35 40.06 40.02
Taiwan-LLM-13B-v2.0-chat 27.74 33.69 27.03 29.43 29.47
Taiwan-LLM-7B-v2.1-chat 25.58 31.76 27.36 27.61 28.08

Inference Performance

In this test, we use the first 700 characters of the web article as the input and ask the model to write the same article again. All inferences run on 2 RTX A6000 GPUs (using vllm, with a tensor-parallel size of 2).

Models ↓ Inference Time (sec) Estimated Max Input Length (Char)
Yi-6B 10.62 5.2k
Breeze-7B-Instruct-v0.1 10.74 11.1k
Breeze-7B-Instruct-64k-v0.1 10.74 88.8k
Qwen-7B 10.86 9.8k
Qwen-14B 18.89 9.8k
Mistral-7B-v0.1 20.48 5.1k
Taiwan-LLM-7B-v2.1-base 26.26 2.2k
Taiwan-LLM-13B-v2.0-base 36.80 2.2k
Yi-34B 43.71 4.5k

Long-context Performance

TBD

Examples

TBD

Use in Transformers

First install direct dependencies:

pip install transformers torch accelerate

If you want faster inference using flash-attention2, you need to install these dependencies:

pip install packaging ninja
pip install flash-attn

Then load the model in transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    model="MediaTek-Research/Breeze-7B-Instruct-v0.1",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    use_flash_attn_2=True # optional
)

The structure of the query template follows that of Mistral-7B-Instruct, as shown below.

<s> SYS_PROMPT   [INST] QUERY1 [/INST] RESPONSE1 [INST] QUERY2 [/INST]

where SYS_PROMPT, QUERY1, RESPONSE1, and QUERY2 can be provided by the user.

The suggested default SYS_PROMPT is

You are a helpful AI assistant built by MediaTek Research. The user you are helping speaks Traditional Chinese and comes from Taiwan.

Citation

@article{breeze7b2024,
  title={},
  author={},
  journal={arXiv},
  year={2024}
}