COCA

This repository hosts the open-source project COCA.

GitHub: https://github.com/qingwang24/COCA

Project Structure

COCA/
├── README.md
├── code/              # Source code for cold start and COCA training.
│   ├── cold_start/    # Cold-start data processing, SFT, and self-play code.
│   ├── coca_train/    # COCA training and iterative GRPO training code.
│   └── readme.md      # Quick pointers to key training scripts.
└── prompts/           # Prompt templates and examples for COCA and evaluation.
    ├── README.md
    ├── coca_prompts.md
    ├── gpt4o_evaluation_prompt.md
    └── gpt4o_output_examples.md

Current Status

• code/: contains the current implementation for cold-start training and COCA iterative training.
• prompts/: contains prompt templates and output examples used by COCA and GPT-4o evaluation.

Code

The main code entry points are summarized in code/readme.md:

• Iterative reasoning: code/cold_start/runner/vllm/self_play.py
• Cold-start SFT: code/cold_start/runner/vllm/train.py
• COCA training: code/coca_train/off_policy_GRPO/grpo_iter_new.py

Prompts

• prompts/README.md: overview of prompt files.
• prompts/gpt4o_evaluation_prompt.md: system prompt for using GPT-4o as an evaluator.
• prompts/gpt4o_output_examples.md: JSON output examples for GPT-4o evaluation.
• prompts/coca_prompts.md: critic and actor prompts used in the COCA method.

Notes

Runtime outputs, logs, checkpoints, and local cache files are intentionally excluded from version control.