Spaces:
Sleeping
Sleeping
Commit History
-
feat: enforce strict (0, 1) score range for Phase 2 and fix UI scroll behavior
Viswanath1293bbd225 -
Fix FastAPI Response typing and mapped dynamic classifier stats
Viswanath129c705daa -
Implement segmented speed control buttons for mails/sec and sync auto-agent rate
Viswanath129efdf547 -
Apply backend formatter updates
Viswanath129e2d8d59 -
Update UI to show Current Task Reward and fix unbounded total_reward observation
Viswanath12948d7965 -
fix: normalize grader with task-based maxima
Viswanath1299f9a567 -
fix: cap progress_ratio at 1.0 to prevent reward exceeding 1.0
Viswanath129- Dynamic emails can cause processed_emails > initial_email_count - This caused progress_ratio > 1.0, inflating progress_bonus beyond 0.15 - Led to step rewards exceeding 1.0 (seen as 1.50 in logs) - Now progress_ratio = min(1.0, processed/initial) to cap bonus at 0.15 - Guarantees step_reward always in [0.0, 1.0]745e84b -
fix: add guard against division by zero in grader() when max_steps=0
Viswanath129- Added check to return 0.0 when max_steps <= 0 - Added max(0.0, ...) guard for efficiency_ratio calculation - Prevents runtime error in edge cases or misconfigured taskse0d8dc3 -
fix: normalize reward function to strict [0.0, 1.0] range per OpenEnv spec
Viswanath129- Fixed critical indentation error in step() method - Replaced additive progress rewards with normalized formula - Removed unbounded +0.5 completion bonus - Replaced grader() with achievement-based scoring: - 60% weight on average action quality - 25% weight on correct action ratio - 15% efficiency bonus (if inbox cleared) - Ensured every step reward is strictly [0.0, 1.0] - Eliminated gradient explosion risk for PPO training - All intermediate calculations now respect normalization constraintefa6c17 -
CRITICAL FIX: Remove stale frontend bundle, fix /step API crash, fix validation handler
Viswanath129c26bbb3 -
FINAL: Min-Max grader, explicit defer handling, all rewards strictly 0.0-1.0
Viswanath1291dab6c5 -
STRICT 0.0-1.0: Hard clamp on all reward paths, no negatives ever, slower demo speed
Viswanath12901f31e5 -
Compliance Phase 2: Final Native 0.0-1.0 Reward Migration
Viswanath129563ed00 -
Compliance Phase 2: Integrated Multimodal HTML State and Enhanced Reward Consistency
Viswanath129cc6e20e -
Compliance Phase 2: Standardized Task Registry and 0.0-1.0 Reward Normalization
Viswanath129343c838 -
Compliance architecture adjustments for OpenEnv Phase 2 waitlist checkout
Viswanath129260ea31 -
Fix OpenEnv architectural compliance
Viswanath129- Remove invalid Environment inheritance from EmailEnv - Fix step endpoint to return proper StepResult tuple format - Update step method to return (observation, reward, done, info) tuple - Fix inference.py to handle new tuple format from step endpoint - Maintain proper reward normalization and OpenEnv compliance - Add health check endpoint for Docker monitoringc71d30c -
Critical reward system fixes
Viswanath129- Fix reward normalization to include completion bonus in calculation - Remove duplicate normalization code in step function - Redirect all debug/info messages to stderr in inference.py - Ensure all rewards are properly normalized to [0.0, 1.0] range - Fix OpenEnv compliance for reward signals263d3e6 -
Complete OpenEnv specification compliance
Viswanath129- Restore full openenv.yaml with tasks, actions, observations, grader - Fix clamp_score functions to allow 0.0 and 1.0 values (remove artificial bounds) - Add reward normalization to ensure per-step rewards in [0.0, 1.0] range - Verify typed models, endpoints, 3 tasks with graders, Dockerfile, README - All requirements now satisfied for OpenEnv hackathon submission62d9e6f
-
fix: clamp ALL scores at every layer - server /step, /grader, env info, Pydantic model
Viswanath129bac365f -
fix: clamp all task scores to strict (0,1) range — never 0.0 or 1.0
Viswanath1293583c75 -
Complete overhaul to fix phase 1: consolidated server package, updated Dockerfile, and fixed imports
Viswanath1293fade06