Spaces:
Running
Running
Commit History
-
refactor: unify data format and add stateful HTTP server
Data Format Changes: - Replace corpus.json/tasks.json with per-seed JSON files - Rename fields: gold_answer→truth, gold_chunk_ids→supporting_items, difficulty→level - Add SupportingItem and DistractorItem models with verification fields - Add sample/ directory with mock data matching generator output format Training Module Updates: - Change difficulty (easy/medium/hard) to level (0/1/2) throughout - Update CurriculumScheduler to use integer levels - Update metrics to use success_rate_by_level dict - Update evaluation to group by level instead of difficulty Server Changes: - Replace stateless openenv server with custom stateful FastAPI server - Environment state now persists across HTTP requests - Add /tasks endpoint to list available tasks Production Data Loading: - Add load_tasks_from_directory() for per-seed file loading - Add load_verified_tasks() and load_tasks_by_level() helpers - Add get_directory_statistics() for production data analysis All 76 tests pass. Co-Authored-By: C…79b7d56