Spaces:
Sleeping
Sleeping
Commit History
-
feat: add TRL/GRPO training support with curriculum learning
garg-tejas- Add training_client.py with LogAnomalyTrainingEnv and CurriculumLogAnomalyEnv - Add training_prompts.py with curated system/user prompts for training - Add train_grpo.py CLI script for GRPO training with vLLM support - Update openenv.yaml with training configuration section - Update pyproject.toml with [training] and [training-vllm] extras - Update server/app.py with concurrent session support (64 max) - Update README.md with training documentation - Add sample LogHub HDFS data for realistic log patterns69126fb
-
feat: winner-inspired improvements - expand reward range, add LogHub data, and hidden state/decoys
garg-tejasPhase 1: Expand Reward Range (-2.0 to +5.0) - Add reward constants to config.py (timeout, wrong type/component penalties, perfect+fast bonus) - Modify grader.py to calculate rewards in expanded range - Update timeout penalty from -0.5 to -2.0 - Perfect + fast answers can now reach +5.0 - Wrong answers incur -0.5 penalties - Provides stronger learning signals for RL training Phase 2: Add Real LogHub Data - Add BGL sample (100 lines from Blue Gene/L supercomputer) - Add OpenStack sample (100 lines from cloud platform) - Add Apache sample (100 lines from web server) - Real data alongside synthetic for more realistic training Phase 3: Add Hidden State / Decoys (POMDP) - Add num_decoys field to DifficultyConfig (easy=0, medium=1, hard=2) - Implement inject_with_decoys() in AnomalyInjector - Decoys are different anomaly type or component, lower intensity - Agent must identify PRIMARY anomaly among multiple suspicious patterns - Grader penalizes decoy matches with -0.3 reward - Transforms f…81b1527