Richard-ZZZZZ/mira-csgo-style-sgfplus-migration-20261012

Mira CSGO map-style SGF+ migration bundle

This bundle is a reproducibility hand-off for the six-style Counter-Strike 2 Dust2 world-model project. It was assembled from the current working tree while the 32-GPU training run continued; it is not a clean checkout of the old Git HEAD.

Contents

• code/mira/: current Mira source, configs, scripts, pixi.toml/pixi.lock, and Git provenance.
• code/h3_pipeline/: the validated Flux-2/H3 first/last-frame and video-editing orchestration subset plus PIPELINE.md.
• data/: three complete P=4 windows (12 player clips) plus ten early p00-only visual samples (three overlap, yielding 19 unique clip directories), source action/state metadata, five edited styles, peer-map cache entries, global map condition metadata, style prompts, and the six-style selection whitelist. This is intentionally a small smoke-test set, not the full training dataset.

The included data/style_six_whitelist.json is provenance from the source Ceph tree and contains source-machine absolute paths; do not use it as the extracted dataset root. The portable smoke loader uses data/source_clips directly and the relative style roots shown in docs/REPRODUCTION.md.

• weights/sgf_plus_step500/: the exact SGF+ continuation model and training_state.pth from checkpoint 500.
• weights/real_score/: real-score teacher model checkpoint used by the SGF+ run.
• weights/stage2_5/: Stage-2.5 six-style warm-start model.
• experiments/h3_qc/: small previously validated H3/Flux QC artifacts.
• docs/: task overview, reproduction notes, experiment evidence, inventory, and original CephFS paths.
• skills/mira-sgfplus-reproduction/SKILL.md: a self-contained hand-off skill for another agent.

Important technical truth

The copied SGF+ files are the Mira SGF+ compatibility adapter on the existing Block-DF Self-Forcing backend. They are not a faithful upstream dual-full-expert implementation: dual_expert=false, dual_full_expert=false, and the backend is sf_rollout. The experiment is useful and currently stable, but a new agent must not claim upstream SGF+ equivalence without implementing and validating the dual-expert/bounded-replay pieces described in docs/KNOWN_LIMITATIONS.md.

Exact continuation artifact

weights/sgf_plus_step500/checkpoint-500/training_state.pth is required for an optimizer/scheduler-exact resume from Step 500. The student model alone is checkpoint.pth. The source run was still active during packaging; its latest observed node-0 log was around Step 966. Do not overwrite the source run or release its machines.

Quick start on a new machine

1. Unpack this archive and set PROJECT_ROOT to the extracted directory.
2. Install the environment using code/mira/pixi.toml and code/mira/pixi.lock (the original run used the Mira pixi default environment).
3. Provide the full six-style dataset separately; the small map-condition and peer-map assets needed by the smoke test are included here.
4. Follow docs/REPRODUCTION.md for the safe, destination-specific resume command and docs/KNOWN_LIMITATIONS.md before changing the algorithm.
5. Validate the sample set with the commands in docs/REPRODUCTION.md before touching a long run.

No Hugging Face token, W&B secret, or other credential is stored anywhere in this bundle.