ZeroEval: A Unified Framework for Evaluating Language Models

ZeroEval is a simple unified framework for evaluating (large) language models on various tasks. This repository aims to evaluate instruction-tuned LLMs for their zero-shot performance on various reasoning tasks such as MMLU and GSM. We evaluate LLMs with a unified setup by controlling the factors such as prompting, sampling, output parsing, etc. In ZeroEval, we perform zero-shot prompting, and instruct LM to output both reasoning and answer in a json-formatted output. We are actively adding new tasks. Contributions are welcome!

• Leaderboard: https://hf.co/spaces/allenai/ZeroEval
• X post

Todo

• Support new tasks (GPPA, AIME, etc.)
• Prefix-prefill for open models such that the parsing is easier
• Add other formatting options (e.g. markup language instead of json, etc.)

Installation

Click to expand

conda create -n zeroeval python=3.10
conda activate zeroeval
# pip install vllm -U # pip install -e vllm 
pip install vllm -U
pip install -r requirements.txt
# export HF_HOME=/path/to/your/custom/cache_dir/ 

Tasks

• MMLU-redux (-d mmlu-redux)

• ZebraLogic (-d zebra-grid)

• CRUX (-d crux)

• MATH (Level 5) (-d math-l5)

• GSM8K (-d gsm)

• More tasks will be added soon. (e.g., ARC, MMLU-Pro, etc.)

Usage

zero_eval_local.sh and zero_eval_api.sh are the two main scripts to run the evaluation.

Examples

• bash zero_eval_local.sh -d mmlu-redux -m meta-llama/Meta-Llama-3-8B-Instruct -p Meta-Llama-3-8B-Instruct -s 4 (Run Llama-3-8B-Instruct with greedy decoding on mmlu-redux)

• bash zero_eval_api.sh -d gsm -f openai -m openai/gpt-4o-mini-2024-07-18 -p gpt-4o-mini-2024-07-18 -s 8 (Run gpt-4o-mini with greedy decoding on gsm)

• bash zero_eval_api.sh -d zebra-grid -f openai -m deepseek-chat -p deepseek-chat -s 8 (Run deepseek-chat via openai style api, with greedy decoding on zebra-grid)

More examples can be found in the scripts folder, e.g., the scripts/_MMLU_redux.md and scripts/_GSM.md files as well as scripts/local/crux.sh.

Arguments

Command Line Arguments

Arguments	Description	Default
-d	DATA_NAME: mmlu-redux, gsm, math-l5, zebra-grid, alpaca_eval, ... (see src/task_configs.py)
-m	model_name
-p	model_pretty_name
-s	number of shards (When -s 1 we'll use all your GPUs for loading the model and running the inference; When -s K, we'll use K GPUs and divide the data into K shards for each GPU to run the inference on a single shard, and merge the results at the end.)	1
-f	engine (vllm by default for zero_eval_local.sh, can be changed to hf; For zero_eval_api.sh, we can use openai, anthropic, ...)	vllm/openai for zero_eval_local/api.sh
-r	run_name (the results will be saved in a sub folder with the run_name when it is specified)	"default"
-t	temperature	0 (greedy decoding)
-o	top_p for nucleus sampling	1.0
-e	repetition penalty	1.0
-b	batch size	4
-x	max_length	4096

Results

🚨 View results on our Leaderboard: https://hf.co/spaces/allenai/ZeroEval

• MMLU-Redux: python src/evaluation/mcqa_eval.py mmlu-redux --> Full results
• GSM/MATH-L5: python src/evaluation/math_eval.py math-l5/gsm --> Full results
• ZebraLogic: python src/evaluation/zebra_grid_eval.py --> Full results and Leaderboard
• CRUX: python src/evaluation/crux_eval.py --> Full results
• All: python src/evaluation/summarize.py --> Full results ⬇️

Citation

If you find ZeroEval useful, please cite it as follows in your publication:

@software{Lin_ZeroEval_A_Unified_2024,
    author = {Lin, Bill Yuchen},
    month = jul,
    title = {{ZeroEval: A Unified Framework for Evaluating Language Models}},
    url = {https://github.com/WildEval/ZeroEval},
    year = {2024}
}

Star History