DATA.md

LowBitMoss research data

Licensed under Apache License 2.0; source model attribution is in THIRD_PARTY_NOTICES.md.

Directory or file	Contents
prompts/calib_prompts.json	General calibration prompt definitions
prompts/calib_text_prompts.json	Text-rich calibration prompt definitions
prompts/eval_prompts.json	The held-out 24-prompt evaluation set
data/calib_512/	Non-distilled teacher sampling captures: prompt text and {x, t, ctx} sample records
data/calib_turbo_768/	Distilled-teacher captures: four conditional sampler states per prompt
data/teacher_out_512.pt, data/teacher_out_512_plus.pt	Non-distilled teacher outputs in sorted capture-file / sample order
data/teacher_out_turbo768.pt	Distilled-teacher outputs in sorted capture-file / sample order
data/enc_cond/	Cached text-conditioning tensors for encoder comparisons
eval/	Generated PNGs, latent tensors, metric JSON, and visual-judge reports

The .pt files use PyTorch's tensor serialization. Use a current PyTorch release and torch.load(path, map_location="cpu", weights_only=True) where supported. Capture samples are Python dictionaries/lists of tensors; the teacher-output files contain tensors ordered to match the capture scripts. Do not reorder captures when pairing them with outputs.

The supplied data/calib_512/ directory contains the expanded 240-prompt pool (4,442 captured samples). Pair it with data/teacher_out_512_plus.pt when running end-to-end tuning. The legacy teacher_out_512.pt belongs to the original 160 general prompts (2,966 samples), so it must not be paired with the expanded directory. For the distilled pool, pair data/calib_turbo_768/ (456 prompts, four samples each) with data/teacher_out_turbo768.pt. If you capture a different pool, regenerate its teacher outputs with the capture files in the same sorted order.

Keep data/ and eval/ under the directory selected by LOWBITMOSS_HOME when using the GitHub scripts. Downloaded model components go in ComfyUI's model directories, as described in the model card. PNG metadata is removed in the public export; pixels and recorded metric values are preserved. Metadata-only privacy redactions, if required, are recorded in the release manifest without exposing original values.

The paper explains sampler settings, held-out subsets, resolution differences, seeds, failed runs, and configuration-specific evidence. Stored judge reports are historical observations; the corrected paper takes precedence when a historical report uses an unsupported interpretation.