Spaces:
Sleeping
Sleeping
Commit History
8 days ago
-
Train and evaluate learned agents; one-click local GPU pipeline
Results on held-out seeds (tasks 1-5), all saved under results/: - Masked PPO: 0.917 from scratch, 0.986 behaviour cloning, 0.993 BC -> PPO. - Qwen3-1.7B (4-bit, LoRA) as a pure operator: 0.200 untrained, 0.515 after a supervised warm start on expert actions, 0.619 after GRPO, trained on one 8 GB RTX 4060 Laptop GPU. Training - train_grpo.py: supervised warm start, then GRPO on decisive states (expert action beats waiting); 8 samples per prompt, micro-batches of 4, VRAM cap and per-step cache clearing so Windows cannot spill into shared memory. - Compact text observation for small models (baseline 0.10 -> 0.20). - run_local_gpu.ps1 / .cmd: baseline eval, training, evaluation of both adapters; eval_llm_local.py gains --label. - Black start pays the remaining in-band bonus on early completion. Docs - README results for PPO and LLMs with analysis, SVG charts built by scripts/make_figures.py (no binary files), CHANGELOG. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.c…76c8761 -
NexusGrid v3: seeded scenarios, physics-based graders, baselines and training code
Environment - DC power flow per island (Kirchhoff B-matrix), primary frequency response, inertia, relay protection (3-tick and instant trips), UFLS and island collapse - Seeded scenario families per task with held-out evaluation seeds (>= 10000); Task 2 draws from a validated contingency catalog split 80/20 train/eval - Attack physics: phantom injection hides real generation loss behind falsified SCADA, forced oscillation grows until damped by a matched counter-signal - Hidden attack details never reach observations (debug_state() for observers) - New action shed_load; out-of-service lines cannot be re-closed - Dense reward = potential-based task progress + frequency/safety terms Grading - Rubrics redesigned so doing nothing or acting randomly scores low - Served-load gate stops "fixing" frequency by cutting customers Agents and tooling - policies.py: no-op, random and an observation-only expert (DC flow checks) - inference.py: pure LLM evaluation by default; fallback is opt-in…8ad4728
-
Add manifest.json endpoint, update dashboard and app
74965f9
-
OpenEnv Hackathon Final Submission
3e4c342