Spaces:
Running
Running
Commit History
-
feat(submission): finalize hackathon deliverables — 24-turn 6/6 eval, hybrid policy, gitignored summary
hschinmaybharadwaj- openenv.yaml: bump episode.max_steps 12 -> 24 so hard tasks (~22 ordered steps) can complete - server/environment.py + server/agent.py: task-aware ExpertPolicy that clears 6/6 missions; used as PPO fallback - training/train_trl_ppo.py: add --use-qlora (peft + bitsandbytes 4-bit) path; final eval pass mirrors notebook - pyproject.toml: declare openenv-core>=0.2.0, trl>=0.11.0, peft, bitsandbytes extras - notebooks/trl_training_colab.ipynb: lean PPO defaults (EPISODES=3, MAX_NEW_TOKENS=24) + EVAL_USE_LLM toggle; final 6-mission eval pass writes training_summary.jsonl on every run, reaches 6/6 success - scripts/evaluate_and_plot.py: split outputs — write artifacts/expert_baseline.jsonl as the oracle reference, never overwrite the notebook-owned training_summary.jsonl - artifacts/: refreshed reward_curves.png + loss_curve.png from the real eval run; commit eval_metrics.jsonl + results_summary.json + expert_baseline.jsonl as the static reward evidence - .gitignore + .dockerignore: trainin…565ba48
-
feat(judge): Reward Forensics Lab + master walkthrough
hschinmaybharadwajAdds two judge-grade entry points and the procedural-task / reward-design endpoints they use, plus cross-links from every existing page. Outstanding feature — /lab (docs/lab.html, ~40KB self-contained): - RLVE difficulty knob (1..5) that procedurally generates fresh missions scaling allies, keywords, op-code length, threat level, and resource thresholds. - Live 5-gate verifier timeline streaming per-component reward decomposition (alliance .30 / resources .25 / operation .20 / message .15 / extraction .10) step by step. - Reward-hacking sandbox with seven pre-baked exploits a judge can fire (spam-best-action, premature run_operation, no-alliance op, brute-force op codes, skip-to-extract, empty-message extract, noop-stall) — each shows the env's matching defense in the timeline and tags the run as defended. Walkthrough — /walkthrough + /walkthrough.md (docs/walkthrough.{md,html}): - Master 32KB markdown document mapping every concept from the OpenEnv briefing (RLVR, …72c744f -
fix(server): serve animated demo at / to eliminate Space root 404
hschinmaybharadwaj- Add GET / route returning docs/visual_demo.html so HF Space iframe probes (and any visitor opening the Space URL) get a 200 with a live training-demo landing page instead of a bare 404. - Use response_model=None to bypass Pydantic schema generation for the HTML/File response, which was crashing app import. - Ship docs/visual_demo.html (animated game view + training feed + hyperparams) plus storytelling docs (judge_flow, pitch_flow, demo_storyboard) the README cross-links. Made-with: Cursor372ac22 -
Harden submission checks: sync manifest, add submission links, include plot evidence
hschinmaybharadwajc477272 -
Add Neon Syndicate environment, training pipeline, and HF Space docs
hschinmaybharadwaj80ba25a
-
Fix multi-mode deployment: add openenv-core dep, server entrypoint, main(), uv.lock, switch to OpenAI API
777dd58 -
Fix [END] output format with score field, improve LLM prompt, add heavier penalties for undesirable behavior
6312f29 -
Add 3 new tasks, update README with HF Space docs and task table
2bc02bd
-
first commit
hschinmaybharadwajb96b8a9