Llama3-8b-chinese-chat-32k

• 📄Paper
• 📚Dataset Download
• ✨GitHub

训练方式

• 使用 NTK-aware 方法扩展上下文长度至 32k

• 以 shenzhi-wang/Llama3-8B-Chinese-Chat 为基础 在 Long-Instruction-with-Paraphrasing 数据集上，使用 QLora 微调 1 epoch。

使用方法

和原版相同

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "yuyijiong/Llama3-8B-Chinese-Chat-32k"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

messages = [
    {"role": "user", "content": "写一首诗吧"},
]

input_ids = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=32768,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

Long-Context Performance

相比原始版本，拥有更强的长上下文能力

LongBench (en)

model	hotpotqa	multifieldqa_en	passage_retrieval_en	qmsum	trec
llama3-8b-chinese-chat	45.88	50.56	68.00	22.52	73.00
llama3-8b-chinese-chat-32k	47.64	49.98	100.00	25.13	75.0

Longbench (zh)

model	dureader	multifieldqa_zh	passage_retrieval_zh	vcsum	lsht
llama3-8b-chinese-chat	29.08	58.4	93.5	14.61	28.25
llama3-8b-chinese-chat-32k	32.31	58.66	82.5	16.15	38.5