RepoQA: Evaluating Long-Context Code Understanding

🏠 Homepage: https://evalplus.github.io/repoqa.html

🚀 Installation

# without vLLM (can run openai, anthropic, and huggingface backends)
pip install --upgrade repoqa
# To enable vLLM
pip install --upgrade "repoqa[vllm]"

pip install --upgrade "git+https://github.com/evalplus/repoqa.git"                 # without vLLM
pip install --upgrade "repoqa[vllm] @ git+https://github.com/evalplus/repoqa@main" # with vLLM

git clone https://github.com/evalplus/repoqa.git
cd repoqa
export PYTHONPATH=$PYTHONPATH:$(pwd)
pip install -r requirements.txt

🏁 Search Needle Function (SNF)

Search Needle Function is the first and base RepoQA task which aims to practice LLMs' ability of long-context code understanding and retrieval. Its corresponding real-life scenario is to perform precise code search from function description.

You can run the SNF evaluation using various backends:

OpenAI Compatible Servers

repoqa.search_needle_function --model "gpt4-turbo" --backend openai
# 💡 If you use openai API compatible server such as vLLM servers:
# repoqa.search_needle_function --base-url "http://localhost:[PORT]/v1" \
#                               --model "Qwen/CodeQwen1.5-7B-Chat" --backend openai

Anthropic Compatible Servers

repoqa.search_needle_function --model "claude-3-haiku-20240307" --backend anthropic

vLLM

repoqa.search_needle_function --model "Qwen/CodeQwen1.5-7B-Chat" --backend vllm

[!Note]

Reference evaluation time:

• Llama3-8B-Instruct: 45 minutes on 2xA6000 (PCIe NVLink)
• Llama3-70B-Instruct: 100 minutes on 4xA100 (PCIe NVLink)

HuggingFace transformers

repoqa.search_needle_function --model "Qwen/CodeQwen1.5-7B-Chat" --backend hf --trust-remote-code

[!Tip]

Installing flash-attn and additionally set --attn-implementation "flash_attention_2" can largely lower the memory requirement.

Google Generative AI API (Gemini)

repoqa.search_needle_function --model "gemini-1.5-pro-latest" --backend google

CLI Usage

• Input:
  • --model: Hugging-Face model ID, such as ise-uiuc/Magicoder-S-DS-6.7B
  • --backend: vllm (default) or openai
  • --base-url: OpenAI API base URL
  • --code-context-size (default: 16384): #tokens (by DeepSeekCoder tokenizer) of repository context
  • --caching (default: True): accelerate subsequent runs by caching preprocessing; --nocaching to disable
  • --max-new-tokens (default: 1024): Maximum #new tokens to generate
  • --system-message (default: None): system message (note it's not supported by some models)
  • --tensor-parallel-size: #GPUS for doing tensor parallelism (only for vLLM)
  • --languages (default: None): List of languages to evaluate (None means all)
  • --result-dir (default: "results"): Directory to save the model outputs and evaluation results
  • --ignore-comments (default: False): During evaluation, ignore groundtruth and model comments
  • --trust-remote-code (default: False): allow remote code (for HuggingFace transformers and vLLM)
  • --attn-implementation (default: None): Use "flash_attention_2" if your HF hits OOM
• Output:
  • results/ntoken_{code-context-size}/{model}.jsonl: Model generated outputs
  • results/ntoken_{code-context-size}/{model}-SCORE.json: Evaluation results

Compute Scores

By default, the repoqa.search_needle_function command will evaluate model outputs and compute scores after text generation. However, you can also separately compute scores using the following command:

repoqa.compute_score --model-output-path={model-output}.jsonl

[!Tip]

• Input: Path to the model generated outputs.
• Output: The evaluation scores would be stored in {model-output}-SCORES.json

📚 Read More

• RepoQA Homepage
• RepoQA Dataset Curation
• RepoQA Development Notes