fix: switch eval to A100-80GB and track completion bonus

chore: switch to A100-80GB for cheaper training runs

feat(v3): hybrid GRPO + SFT training with expert beam search

fix(v3): confidence-gated epsilon to avoid destroying JSON structure

feat(v3): epsilon-greedy exploration to guarantee GRPO diversity

feat(v3): add top-k sampling for forced token diversity during exploration

feat(v3): expose noise_scale and temperature as Modal CLI args

fix: unbuffered stdout for real-time Modal training logs

perf(v3): batch reference log-probs to reduce adapter toggling overhead

fix(v3): prevent collapse from noise-amplified negative gradients

feat(v3): exploration noise processor for GRPO output diversity

fix(v3): prompt perturbation to break deterministic outputs

fix(v3): correct default lr in modal entrypoint to match train_v3.py

fix(v3): guarantee reward variance with num_per_task and higher temp

fix(v3): normalize KL per token to prevent over-skipping

fix(v3): prevent catastrophic forgetting with KL constraints

fix(v3): improve training diagnostics and autocast deprecation

fix(v3): include multi-step tasks from curriculum start

fix(v3): proper reference model handling for KL divergence

feat(v3): custom multi-step GRPO training loop with GiGPO

feat(v3): add rollout logic and fix target_crease_edges bug

docs: add V3 sequential RL training architecture and implementation plan

feat(v3): implement core building blocks for multi-step RL training

fix: default port to 7860 for HuggingFace Spaces health check

feat: V2 frontend — episode player UI

chore: track binary assets with git-lfs

feat(v2): implement multi-step environment with PaperState and per-step rewards

feat(v2): filter None fields from OrigamiAction in client step payload

feat(v2): update train_grpo.py for step-level prompts and per_step_reward

feat(v2): update OrigamiAction/Observation/State for multi-step mode

feat(v2): add step_reward.py — per-step Kawasaki/Maekawa/coverage reward

feat(v2): add max_folds to tasks + waterbomb_base + map_fold tasks

feat(v2): add extract_crease_json and valid_crease reward to training/reward.py

feat(v2): port PaperState to origami_server/engine/paper_state.py

chore(v2): add shapely dependency for PaperState intersection detection

feat(v2): port CreaseGraph to origami_server/engine/graph.py

docs: add V2 handoff and plan documents

feat: add hackathon demo script and result charts

feat: add Modal eval script for GRPO checkpoints

feat: update origami viewer UI

feat: expand reward pipeline and GRPO training config

chore: ignore pycache, outputs, and trainer state

chore: add HF Spaces frontmatter to README

revert: restore train_origami.ipynb to Prasanna's exact version

chore: restore Prasanna's notebook, add Modal launch section

feat: Railway deployment + multi-task GRPO + Modal B200 training

Update training notebook

Add MAX_CONCURRENT_ENVS, sync latest changes

Update viewer

Add GRPO training notebook + Dockerfile for cloud training (#1)