LLM Collaboration - Code Completion

This repo provides the extended environments for CoMLRL.

Installation

Install CoMLRL:

pip install comlrl
# Install PyTorch compatible with your device

Or via conda-forge:

conda install -c conda-forge comlrl
# Install PyTorch compatible with your device

Dataset: ClassEval

• Contents: each sample includes a class skeleton, method stubs (with docstrings or pass), and canonical hidden tests.
• Splitting: train/train_grpo.py performs a stable random train/eval split (default 80/20) on top of the Hugging Face dataset; seeds are optionally randomized via fixed_seed=false.
• Prompting: prompts include the sanitized class skeleton, explicit method names per agent, and any collaboration instructions.
• Testing: reward code merges agent completions back into the skeleton and runs the provided unit tests inside a temporary directory to isolate state.

Settings

Key sections in configs/config.yaml:

• model: base checkpoint (Qwen/Qwen2.5-Coder-3B-Instruct by default), tokenizer/model kwargs, and device mapping.
• data: dataset name and split ratio; customize when experimenting with different ClassEval sub-splits or local mirrors.
• collab: choose ONE or TAKE_JOB and set num_agents (for TAKE_JOB).
• external: determines the feedback mode. token_report summarizes syntax/tests at each turn; other modes replicate the options documented in the code-generation README (plain, level_feedback, group_feedback, personal_feedback, personal_detailed_feedback, passed, level_passed).
• trainer: forwarded to comlrl.trainers.magrpo.MAGRPOTrainer. Includes sampling settings (num_generations, num_turns, temperature/top_p), optimization hyperparameters, and IO controls (output_dir, logging_steps, etc.).
• reward_processor: optional extra scaling/shift before the reward hits the trainer.
• output: persistence knobs (save final model, keep tmp dirs); environment variables such as CLASSEVAL_TMP_BASE are derived from this section to colocate temp files per job.

Rewards, Logging, and Evaluation

• rewards/CE_reward.py computes structured rewards:
  • lv1: coverage of unique methods completed.
  • lv2: penalizes under/over-allocation of total method picks.
  • lv3: balance term encouraging an even workload across agents.
  • lv4/lv5: syntax + unit-test bonuses (reported for analysis; syntax/test failures short-circuit the run where applicable).
• Tests execute inside per-sample temporary directories to avoid polluted state and are automatically truncated on timeout.
• Loggers are inherited from CoMLRL. Enable Weights & Biases by filling wandb.entity or disable it for offline debugging.