Fetching metadata from the HF Docker repository...
-
Delete baseline/__pycache__
-
Fix Env: Change Model
-
Expose data_lag_days in CityObservation; remove fragile num_districts==6 inference
-
Reduce rollouts easy=2 medium=3 hard=3 — target ~15min runtime with buffer for judge overhead
-
Fix 8 issues spotted in code review: Dockerfile path, task descriptions, duplicate function, fallback score consistency, twin constants comment, has_data_lag docs
-
Remove Validation.py
-
Cut easy rollouts 3→2 to stay under 20min; update GRPO scores; fill config/tasks.yaml
-
Fix All: Cleaned All Files
-
Fix Readme: Corrected Grader Percentages
-
Easy rollouts 2→3 for better score stability, still under 20min
-
Per-task rollouts: easy=2, medium=4, hard=4 — reduces runtime to ~16min
-
Reduce max_tokens 120→60 — agent uses ~32 tokens, no content affected
-
Fix Inference,Evaluator: Increased Rollouts to 4
-
Fix Inference,Env: Change Model
-
Fix memory threshold, pre-compute lag estimates in hard prompt
-
Fix memory threshold, pre-compute lag estimates in hard prompt
-
Fix Inference,Env: Change Model
-
Fix Evaluator: Decrease Rollouts to 3
-
Fix evaluator:Increase number of rollouts to 8 for better performance.
-
Improve agent: fix strategy rule, chain-of-thought, 8 rollouts, better memory
-
Fix treatment/recovery calibration, hospital breach reward, efficiency grader, recalibrate seeds
-
Fix realism: natural recovery in spread, hospital breach at 10%, linear spillover, correct prompt
-
Increase rollouts to 5 for better GRPO learning convergence
-
Swap grader weights: hospital=45% primary constraint, containment=30%, calibrate environment
-
Lower spread rate to 0.09, fix medium seeds, teach focused allocation strategy
-
Raise treatment reduction, lower medium/hard seeds, focus prompt strategy
-
Fix resource replenishment, grader efficiency, seed calibration, prompt strategy
-
Add /grade endpoint, use real grader scores in evaluator
-
Add README.md
-
Initial submission: Cascade Containment
