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_magrpo.py loads explicit HF slices from dataset.train_split and dataset.eval_split (e.g., test[:50] and test[50:]).
• Subsetting: if a split name is missing (e.g., ClassEval only has test), the loader falls back to the first available split before slicing.
• Prompting: prompts include the sanitized class skeleton plus per-agent method assignments. The default strategy assigns 1-parameter methods to agent 0 and all other methods to agent 1.
• 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/magrpo_classeval_config.yaml:

• model: base checkpoint (Qwen/Qwen2.5-Coder-3B-Instruct by default), tokenizer/model kwargs, and device mapping.
• dataset: dataset name and split strings (train_split, eval_split) for ClassEval sub-slices or local mirrors.
• external: feedback configuration (use code_feedback for syntax/test diagnostics).
• magrpo: forwarded to comlrl.trainers.magrpo.MAGRPOTrainer. Includes collaboration (num_agents, param-count assignment), sampling settings (num_generations, num_turns, temperature/top_p), rollout buffering (rollout_buffer_size), optimization hyperparameters, and IO controls.
• reward_processor: optional post-processing for rewards (scale, shift).
• output: persistence knobs (save final model, output paths, verbose debug prints).

Rewards, Logging, and Evaluation

• rewards/CE_reward.py computes structured rewards:
  • lv1: syntax score proportional to valid method outputs (range [0, 2]).
  • lv2: unit-test bonus based on pass rate (passed/total), scaled to [0, 4].
  • lv3: overlap penalty normalized by total methods (range [-1, 0]).
  • reward shift: optional post-processing shift via reward_processor.shift.
• 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.