Spaces:
Running
Running
Commit History
-
Fix API reward clamp to (0.001, 0.999) and update README Running
- app.py: align API-layer reward clamp to 0.001/0.999 - README: update reward table to reflect all-positive clamped rewards Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1bb11d9 -
Enforce strict (0.001, 0.999) bounds on ALL rewards and scores
- rewards.py: fix early returns for invalid/timeout that skipped clamp - env_core.py: align _REWARD_MIN/MAX to 0.001/0.999 - graders.py: align _SCORE_MIN/MAX to 0.001/0.999 - inference.py: align _clamp01 to 0.001/0.999 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>c314a65 -
Set minimum shaped reward to 0.001 (strict >0)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>3948a09 -
Clamp shaped rewards to non-negative values
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>d5dbfe8 -
Fix terminal reward to be grader score only (strict 0-1 range)
When timeout occurs, the reward was: step_reward + timeout_penalty + grader_score which could be negative and fail the validator's (0,1) score check. Now terminal reward = grader_score only, always in (0.001, 0.999). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>5961585 -
Fix score bounds to (0.001, 0.999) and use HF Router defaults
- Grader _clip bounds: 0.001 to 0.999 (safe with .3f format) - inference.py score format: .3f (matches sample) - inference.py defaults: HF Router + Qwen2.5-72B (validator provides HF_TOKEN) - env_core fallback score: 0.001 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>373c99b -
Tighten score bounds to (0.000001, 0.999999) for strict validation
Validator requires scores strictly between 0 and 1 (exclusive). Widened decimal precision in log output to 6 places to avoid rounding 0.000001 to 0.000 which could be read as 0.0. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>6f37fb0 -
Fix grader scores to be strictly within (0, 1) range
Phase 2 validation requires scores strictly between 0 and 1 (not 0.0 or 1.0). Updated _clip in graders.py to clamp to [0.01, 0.99], fixed env_core.py fallback, and updated inference.py _clamp01 accordingly. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>c5b547b -
Version 3: add trained model checkpoints
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>ab786b3 -
Version 3
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>f0ef01d
-
fix: monorepo
e543908
