MiuLab / TMLU

Taiwanese Mandarin Language Modeling

TMLU Eval

🤗 Dataset • 📃 Paper

For open source models, we support

• Hugging-Face as backend
  • Use the probabilities of the option codes as prediction
• vLLM as backend
  • Parse the model generation to extract the prediction

For proprietary models, we support

• OpenAI API
  • Including custom API server which support openai-python
• Anthropic API
• Google API

Environment Setup

To set up the environment for TMLU Eval, follow these steps:

1. Create a new conda environment:

   conda create -n tmlu python=3.10 -y

2. Activate the environment:

   conda activate tmlu

3. Install PyTorch with CUDA support:

   conda install pytorch pytorch-cuda=12.1 -c pytorch -c nvidia -y

4. Install the required dependencies:

   pip install -r requirements.txt

Hugging Face

Use the probabilities of the option codes as prediction.

Command:

$python3 tmlu_eval.py \
	--backend hf \
	--model [Model name on huggingface or path] \
	--dtype [torch type for the model, default will follow the model config] \
	--subsets [Choose subsets of TMLU (names splited by comma) for evalutaion. Default is 'ALL'] \
	--log_dir [Directory for saving evaluation log] \
	--few_shot_num [The number for few shot example. Range: [0, 5]. Default is 5.]

Example:

$python3 tmlu_eval.py \
	--backend hf \
	--model yentinglin/Taiwan-LLM-7B-v2.1-chat \
	--dtype float16 \
	--temperature 0.0 \
	--subsets AST_chinese,AST_mathematics \
	--log_dir log/prob_based/Taiwan-LLM-7B-v2.1-chat \
	--few_shot_num 5

vLLM

Parse the model generation to extract the prediction.

Command:

$python3 tmlu_eval.py \
	--backend vllm \
	--model [Model name on huggingface or path] \
	--dtype [torch type for the model, default will follow the model config] \
	--temperature [temperature for generation] \
	--max_tokens [max new tokens to generate] \
	--subsets [Choose subsets of TMLU (names splited by comma) for evalutaion. Default is 'ALL']  \
	--tensor_parallel_size [Tensor parallel size for vLLM] \
	--log_dir [Directory for saving evaluation log] \
	--few_shot_num [The number for few shot example. Range: [0, 5]. Default is 5.] \
	--cot [Use CoT evaluation.]

Example:

$python3 tmlu_eval.py \
	--backend vllm \
	--model yentinglin/Taiwan-LLM-7B-v2.1-chat \
	--dtype float16 \
	--temperature 0.0 \
	--max_tokens 128 \
	--subsets AST_chinese,AST_mathematics \
	--tensor_parallel_size 1 \
	--log_dir log/gen_based/Taiwan-LLM-7B-v2.1-chat \
	--few_shot_num 5 \
	--cot

Custom API model (use openai-python for querying)

Parse the model generation to extract the prediction.

Command:

$python3 tmlu_eval.py \
	--backend custom_api \
	--model [Model name] \
	--base_url [base url for the API]
	--temperature [temperature for generation] \
	--max_tokens [max new tokens to generate] \
	--subsets [Choose subsets of TMLU (names splited by comma) for evalutaion. Default is 'ALL'] \
	--log_dir [Directory for saving evaluation log] \
	--few_shot_num [The number for few shot example. Range: [0, 5]. Default is 5.] \
	--cot [Use CoT evaluation.]

Example:

$python3 tmlu_eval.py \
	--backend custom_api \
	--model yentinglin/Taiwan-LLM-MoE-alpha \
	--base_url http://127.0.0.1:8888/v1
	--temperature 0.0 \
	--max_tokens 128 \
	--subsets AST_chinese,AST_mathematics \
	--log_dir log/gen_based/Taiwan-LLM-MoE-alpha

OpenAI

Parse the model generation to extract the prediction.

before start

Set environment variable OPENAI_API_KEY.

Command:

$python3 tmlu_eval.py \
	--backend openai \
	--model [Model name] \
	--temperature [temperature for generation] \
	--max_tokens [max new tokens to generate] \
	--subsets [Choose subsets of TMLU (names splited by comma) for evalutaion. Default is 'ALL'] \
	--log_dir [Directory for saving evaluation log] \
	--few_shot_num [The number for few shot example. Range: [0, 5]. Default is 5.] \
	--cot [Use CoT evaluation.]

Example:

$python3 tmlu_eval.py \
	--backend openai \
	--model gpt-4-1106-preview \
	--temperature 0.0 \
	--max_tokens 128 \
	--subsets AST_chinese,AST_mathematics \
	--log_dir log/gen_based/gpt-4-1106-preview

Anthropic

Parse the model generation to extract the prediction.

before start

Set environment variableANTHROPIC_API_KEY.

Command:

$python3 tmlu_eval.py \
	--backend anthropic \
	--model [Model name] \
	--temperature [temperature for generation] \
	--max_tokens [max new tokens to generate] \
	--subsets [Choose subsets of TMLU (names splited by comma) for evalutaion. Default is 'ALL'] \
	--log_dir [Directory for saving evaluation log] \
	--few_shot_num [The number for few shot example. Range: [0, 5]. Default is 5.] \
	--cot [Use CoT evaluation.]

Example:

$python3 tmlu_eval.py \
	--backend anthropic \
	--model claude-2.0 \
	--temperature 0.0 \
	--max_tokens 128 \
	--subsets AST_chinese,AST_mathematics \
	--log_dir log/gen_based/claude-2.0

Gemini-pro

Parse the model generation to extract the prediction.

before start

Set environment variableGOOGLE_API_KEY.

Command:

$python3 tmlu_eval.py \
	--backend google \
	--model [Model name] \
	--temperature [temperature for generation] \
	--max_tokens [max new tokens to generate] \
	--subsets [Choose subsets of TMLU (names splited by comma) for evalutaion. Default is 'ALL'] \
	--log_dir [Directory for saving evaluation log] \
	--few_shot_num [The number for few shot example. Range: [0, 5]. Default is 5.] \
	--cot [Use CoT evaluation.]

Example:

$python3 tmlu_eval.py \
	--backend google \
	--model gemini-pro \
	--temperature 0.0 \
	--max_tokens 128 \
	--subsets AST_chinese,AST_mathematics \
	--log_dir log/gen_based/gemini-pro

Citation

If you use TMLU in your research, please cite the following paper:

@article{DBLP:journals/corr/abs-2403-20180,
  author       = {Po{-}Heng Chen and
                  Sijia Cheng and
                  Wei{-}Lin Chen and
                  Yen{-}Ting Lin and
                  Yun{-}Nung Chen},
  title        = {Measuring Taiwanese Mandarin Language Understanding},
  journal      = {CoRR},
  volume       = {abs/2403.20180},
  year         = {2024},
  url          = {https://doi.org/10.48550/arXiv.2403.20180},
  doi          = {10.48550/ARXIV.2403.20180},
  eprinttype    = {arXiv},
  eprint       = {2403.20180},
  timestamp    = {Wed, 10 Apr 2024 17:37:45 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2403-20180.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}