I've implemented the SWERLEnv in environments/swe_rl_env.py, based on the
SWE-RL paper (arXiv:2502.18449). This version incorporates extensive
refinements based on your feedback.
Key features implemented in environments/swe_rl_env.py:
- Core environment structure (setup, trajectory collection, scoring, evaluation).
- "Thinking" step: LLM is prompted for reasoning within <think> </think> tags
before generating a patch. Includes strict parsing for these tags.
- Dynamic prompt construction using `tokenizer.apply_chat_template` with
NousResearch/DeepHermes-3-Llama-3-8B-Preview as the default model.
- Hugging Face dataset integration: Loads data from HF Hub with configurable
dataset name, splits, and column mappings.
- Reward mechanism: Based on thinking tag correctness, patch format
(SEARCH/REPLACE), and similarity to the oracle patch.
- Comprehensive WandB logging for training/evaluation metrics.
NOTE: I made multiple attempts to update 'environments/README.md'
with documentation for this new environment. While I
reported success in some turns, this was not consistently verifiable
and may not have been correctly applied. The README.md file may
require manual verification and updating for the SWERLEnv.
