Fetching metadata from the HF Docker repository...
-
Final Hardening: Add 'score=' to [END] log and implement stderr reward auditing
-
Harden reward boundaries: Shift from [0.01, 0.99] to safer [0.05, 0.95] to avoid potential validator threshold issues
-
Fix stdout formatter rounding rewards to 1.00 in Edge Cases
-
Refactor error sanitization
-
Add heuristic fallback agent to inference.py to handle LLM API failures gracefully
-
Refactor: Enforce strict float boundaries and clean up formatting across grader, API, and tests
-
Add test validator script and ensure all changes are synced
-
Fix /step to accept both wrapped and unwrapped action formats from OpenEnv validator
-
Round reward scores to 1 decimal place (0.1-0.9 range, strictly valid)
-
CRITICAL FIX: Clamp reward floats at API boundary in /step endpoint to prevent exact 0.0/1.0 from reaching OpenEnv validator
-
fix(env): terminate episodes on mitigation to restrict task episodic returns under 1.0 bounds
-
fix(schema): strict bounds ge=0.01 and le=0.99 for OpenEnv validation
-
Push pending environment.py reward boundary bounds
-
Fix edge-case explicitly emitting 0.00 rewards on LLM failure in inference.py
-
Fix OpenAPI schema definitions to explicitly exclude 0.0 and 1.0 (gt=0.0, lt=1.0) and enforce bounds in object instantiation
-
Fix reward boundaries to strictly (0, 1) per OpenEnv validator
-
Fix ASGI app import path in Dockerfile and use pyproject.toml
-
Compliance with OpenEnv v0.2.0 multi-mode deployment requirements
-
Add pyproject.toml to satisfy openenv validation
-
Completely bypass Pydantic body validation for /reset and /step to handle raw null/empty strings
-
Fix POST /reset body requirement for openenv validator
-
Add app.py and YAML configuration
-
Add HuggingFace space YAML configuration to README
-
Initial commit for AI SOC Gym Hackathon submission
