BlueSourceJY/shotpilot-stage1-evaluations

ShotPilot Stage 1 — evaluation rollouts and scores

This dataset holds the evaluation results of every scheduled evaluation for two Stage 1 training runs of ShotPilot. ShotPilot is a chunk-level autoregressive, multi-shot text-to-video generator fine-tuned from Wan2.1-T2V-1.3B.

Folder	Run	Weights evaluated	Steps
original/	Original recipe: LR 1e-5, no EMA, clean history	raw	0, 200, …, 4200 (22 evaluations)
recipe_b/	Recipe B, forked from step 1600: LR 2e-6, EMA 0.999, history noise σ ~ U[0, 0.2)	EMA	1800, 2000, …, 2600 (5 evaluations)

The Recipe B checkpoints are in BlueSourceJY/shotpilot-stage1-recipe-b. Training was still running when this snapshot was uploaded on 2026-10-12.

Evaluation suite

Every evaluation uses the same frozen suite (suite.json; protocol generated-rollouts-v2):

• 64 validation clips are generated autoregressively, chunk by chunk, from the enriched global prompt and the ground-truth action/time sequence. No real pixels enter generation: the history is the model's own generated output.
• 32 paired interventions re-generate one target chunk with the action flipped (CUT ↔ CONTINUE). The generated prefix, noise, text and time are kept the same.
• Sampling: 50-step FlowUniPC, CFG 5.0, shift 5.0, base seed 2026091101, 16 fps, 3 latent frames per chunk.

The prompts (clips-prompts.json, shots-prompts.json) are LLM-enriched descriptions of held-out validation clips.

Metrics (metrics.csv)

There is one row per run × step.

Column	Meaning
aesthetic_quality, imaging_quality	VBench-derived, mean over the 64 clips
subject_consistency, motion_smoothness, dynamic_degree	VBench-derived, computed within shots and then averaged per clip
paired_latent_rmse	Latent RMSE between the CUT and CONTINUE branches of each pair (action response)
paired_first_frame_l1_same_decoder	First-frame L1 between the branches, decoded with the same decoder context
cut_boundary_l1, continue_boundary_l1	Pixel L1 across natural CUT and CONTINUE chunk boundaries
paired_cut_minus_continue_boundary_l1, paired_cut_boundary_larger_fraction	Boundary gap within each pair, and the fraction of pairs where the CUT boundary is larger

Caveats:

• This is a custom validation subset, not the official VBench total.
• Dynamic degree is a collapse diagnostic.
• Across checkpoints, motion smoothness is almost a linear readout of motion amplitude, so compare checkpoints at matched motion.

Layout

metrics.csv                              # all runs × steps, summary metrics
<run>/evaluation-history.jsonl           # orchestrator's record of each evaluation
<run>/latest-evaluation.json
<run>/checkpoint-XXXXXXX/
  export.log, rollouts.log, quality.log  # logs of export, generation and scoring
  export/_SUCCESS.json                   # identity/hash of the exported weights (weights not included)
  rollouts/
    suite.json, summary.json             # frozen suite; all scores including per-clip lists
    clips/rowNNNNNN.mp4                  # 64 full multi-shot episodes
    shots/rowNNNNNN-shotKK.mp4           # the same episodes split per shot (input to the within-shot metrics)
    pairs/rowNNNNNN-tT-aA.mp4            # paired intervention: A=1 CUT, A=2 CONTINUE at target chunk T
    clips-prompts.json, shots-prompts.json
    records-rank*.json, rollout-rank*.json  # per-rank generation records, pair diagnostics, timings
    evaluation-complete.json
    vbench/<dimension>/                  # raw VBench results, evidence, full info and score logs
original/figures/stage1_metrics_0_4200/  # metric-curve figure for the original run

Not included: the exported inference weights (export/model.pt), the raw rollout latents, and the ROCm/MIOpen kernel caches. Absolute paths inside JSON and log files refer to the original training cluster.

License

CC BY-NC 4.0. The videos are model outputs, and the model was fine-tuned on non-commercial research video data. Use them for non-commercial research only.