Drop unsloth/vllm dependency chain to eliminate huggingface-hub version
conflicts. Use AutoModelForCausalLM + BitsAndBytesConfig + LoraConfig
directly. Flash attention falls back to SDPA if flash-attn is missing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Renders an MP4 showing the 3D stellarator boundary morphing as the
agent adjusts parameters, with live metrics overlay, reward ticker,
and rotating camera. Uses matplotlib animation with parameter
interpolation between steps for smooth transitions.
Usage: uv run python training/demo_animation.py --seed 0 --output demo.mp4
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add auto-submit to run_episode_with_actions so short non-terminal
plans always include terminal reward terms
- Replace `float(obs.reward or 0.0)` with explicit None check to
match notebook semantics
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
fix: robust JSON array extraction and notebook GRPO fixes
Replace greedy regex with balanced-bracket parser that skips prose
fragments like [draft] and only returns candidates that decode as
valid JSON arrays. Also updates the notebook with:
- single environment_reward_fn (drop redundant format_reward_fn)
- auto-submit for short non-terminal plans so terminal reward fires
- same robust _extract_json_array parser as the packaged module
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
feat: add reward and verifier monitoring telemetry
- Add HTML landing page at / showing task, constraints, and API docs
- Revert notebook to format + environment reward (env already decomposes
feasibility and objective internally via _compute_reward)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
docs: clarify notebook surfaces and OpenEnv guidance
- Add apt-get build deps for constellaration on Colab
- Replace FusionLabClient with requests for robust HF Space demo
- Update README with HF Space and notebook links, mark deployment complete
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
feat: add llm rollout contract and simplify ppo smoke
booz-xform (constellaration dep) needs cmake, ninja, g++, gfortran,
and libnetcdf-dev to build from source. Verified locally.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
fix: quote pip version specs in Dockerfile, add cpu-upgrade
Shell interprets >= as redirect without quotes. Also add
ENABLE_WEB_INTERFACE=true and libgomp1 for runtime.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add root-level re-export files (__init__.py, client.py, models.py)
for OpenEnv packaging convention
- Switch Dockerfile base from openenv-base to python:3.12-slim for
reliable HF Space builds
- Add Colab-ready GRPO training notebook using Unsloth + TRL
with environment reward functions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move to docs/archive/ since it is a tactical checklist, not a planning
SSOT. Fix training-optimism, align fallback language to SSOT section 10,
split training milestones, and soften reward branch claim.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
fix: resolve high-fi submit blocker by switching to from_boundary_resolution preset
The high_fidelity VMEC preset forces minimum 10 modes, which doesn't
converge on our mpol=3/ntor=3 boundaries. Switched submit evaluation
to the from_boundary_resolution preset, which adapts resolution to
the boundary's actual Fourier content and converges reliably (~4s).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Run 5 scripted episodes across all 3 seeds exercising 12/13 reward
branches. Key findings: cross-fidelity gap confirmed (low-fi feasible
state crashes at high-fi submit), elongation crash pocket at ~1.25-1.30,
and feasibility boundary is multi-parametric.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Validates the repaired 4-knob parameter ranges, deltas, and reset seeds
against 396 low-fidelity evaluations (81 broad + 315 targeted). Key finding:
rotational_transform upper bound should tighten from 1.9 to 1.8 to eliminate
the dead crash zone. Current seeds and deltas validated as functional.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>