Spaces:
Sleeping
Sleeping
Commit History
-
Rewrite notebook to follow Unsloth GRPO pattern (trl==0.22.2)
- Fix llm_blender error: replace trl>=0.16 with pinned trl==0.22.2 via uv - Split reward into formatting_reward_func + environment_reward_func - Upgrade GRPOConfig: DR-GRPO loss, cosine LR, adamw_8bit, importance sampling - Add max_prompt_length, max_grad_norm, log_completions, gradient_accumulation_steps Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a9a44e8 -
Fix missing mergekit dependency in Colab install cell
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>fa5485d -
Merge remote changes and fix parse_model_action for GET payloads
Merge remote inference.py (structured logging, MAX_TASK_REWARD, ENV_URL) while keeping the parse_model_action fix: GET requests treat everything after the method as the path so UNION payloads with spaces are not silently truncated into the body field. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>525b2ba -
Add GRPO training notebook and dense reward redesign for Round 2
- Redesign reward function with dense step-level signals, hint field population, recon rewards, efficiency bonus, and seeded milestone tracking - Fix repeat detection to use (key, status) pairs so re-visiting an endpoint after gaining new privileges is not penalized - Fix parse_action in inference.py and training.py: GET requests with spaces in query strings (e.g. UNION payloads) were silently truncated - Add training.py: GRPO training script using Unsloth + TRL - Add bughunter_grpo_training.ipynb: Colab-ready notebook for Gemma 4 GRPO training with before/after evaluation and reward curve plotting Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>26e15b9
-
Fix score formula: proportional to MAX_TASK_REWARD, clamped to (0.001, 0.999)
2bdfabc -
Fix inference.py: correct stdout format and default ENV_URL to HF Space
f1d1d6c -
Add uv.lock
747c0aa -
Move Dockerfile to root for HF Spaces
fe54bd1 -
Initial submission: Bug Hunter Env
65ac9e6 -
initial commit
Verified signature527bbfb