Breeze-7B-Base-v0_1

Model Card for MediaTek Research Breeze-7B-Base-v0_1

MediaTek Research Breeze-7B (hereinafter referred to as Breeze-7B) is a language model family that builds on top of Mistral-7B, specifically intended for Traditional Chinese use.

Breeze-7B-Base is the base model for the Breeze-7B series. It is suitable for use if you have substantial fine-tuning data to tune it for your specific use case.

Breeze-7B-Instruct derives from the base model Breeze-7B-Base, making the resulting model amenable to be used as-is for commonly seen tasks.

Breeze-7B-Instruct-64k is a slightly modified version of Breeze-7B-Instruct to enable a 64k-token context length. Roughly speaking, that is equivalent to 88k Traditional Chinese characters.

Update (Feb. 21st, 2024): Breeze-7B-Instruct-64k-v0_1 has been temporarily removed from Hugging Face due to its actual performance in long context tests not meeting expectations.

Update (Mar. 7th, 2024): The current release version of Breeze-7B is v1.0. See Breeze-7B-Base-v1_0.

Practicality-wise:

• Breeze-7B-Base expands the original vocabulary with additional 30,000 Traditional Chinese tokens. With the expanded vocabulary, everything else being equal, Breeze-7B operates at twice the inference speed for Traditional Chinese to Mistral-7B and Llama 7B. [See Inference Performance.]
• Breeze-7B-Instruct can be used as is for common tasks such as Q&A, RAG, multi-round chat, and summarization.
• In particular, Breeze-7B-Instruct-64k can perform tasks at a document level, not a chapter level.

Performance-wise:

• Breeze-7B-Instruct demonstrates impressive performance in benchmarks for Traditional Chinese and English, when compared to similar sized open-source contemporaries such as Taiwan-LLM-7B/13B-chat, QWen-7B-Chat, and Yi-6B-Chat. [See Chat Model 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-token context length
• Breeze-7B-Instruct-v0_1
  • Expanding the vocabulary dictionary size from 32k to 62k to better support Traditional Chinese
  • 8k-token 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-token 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
  • Finetuned from: MediaTek-Research/Breeze-7B-Base-v0_1
  • Model type: Causal decoder-only transformer language model
  • Language: English and Traditional Chinese (zh-tw)
• Breeze-7B-Instruct-64k-v0_1
  • Finetuned from: MediaTek-Research/Breeze-7B-Instruct-v0_1
  • Model type: Causal decoder-only transformer language model
  • Language: English and Traditional Chinese (zh-tw)

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. All choice problems adapt the selection by the log-likelihood.

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

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. All choice problems adapt the selection by the log-likelihood. We use the code revised from fastchat llm_judge (GPT4 as judge) to evaluate MT-Bench-tw and MT-Bench.

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

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).

Long-context Performance

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(
    "MediaTek-Research/Breeze-7B-Base-v0_1",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2" # optional
)
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("MediaTek-Research/Breeze-7B-Base-v0_1")
tokenizer.tokenize("你好，我可以幫助您解決各種問題、提供資訊和協助您完成許多不同的任務。例如：回答技術問題、提供建議、翻譯文字、尋找資料或協助您安排行程等。請告訴我如何能幫助您。")
# Tokenized results
# ['▁', '你好', '，', '我', '可以', '幫助', '您', '解決', '各種', '問題', '、', '提供', '資訊', '和', '協助', '您', '完成', '許多', '不同', '的', '任務', '。', '例如', '：', '回答', '技術', '問題', '、', '提供', '建議', '、', '翻譯', '文字', '、', '尋找', '資料', '或', '協助', '您', '安排', '行程', '等', '。', '請', '告訴', '我', '如何', '能', '幫助', '您', '。']

Citation

@article{MediaTek-Research2024breeze7b,
      title={Breeze-7B Technical Report}, 
      author={Chan-Jan Hsu and Chang-Le Liu and Feng-Ting Liao and Po-Chun Hsu and Yi-Chang Chen and Da-Shan Shiu},
      year={2024},
      eprint={2403.02712},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}