Skip to content

feat(corpus): add resumable captioned corpus-v2 builder - #15

Merged
vex0209-bt merged 1 commit into
mainfrom
feat/corpus-v2-pilot
Jul 24, 2026
Merged

feat(corpus): add resumable captioned corpus-v2 builder#15
vex0209-bt merged 1 commit into
mainfrom
feat/corpus-v2-pilot

Conversation

@vex0209-bt

Copy link
Copy Markdown
Contributor

Summary

  • add a deterministic, resumable corpus-v2 builder for five-second TI2V truth clips and lossless frame-zero PNGs
  • caption prepared clips with a pinned Qwen3-VL revision using one persistent worker per selected GPU
  • enforce motion-stratified human QA before verified, content-addressed Hippius publication
  • generate digest-anchored manifest shards and publish the immutable root last
  • document the offline build, review, publish, and staged scaling workflow

Safety boundaries

  • leaves chain.toml and the active corpus-v1 evaluator unchanged
  • refuses partial manifest roots and verifies uploaded artifact size plus SHA-256 metadata
  • requires evaluator-v2 support and calibration before a corpus-v2 root may be pinned into consensus
  • preserves restart safety through a SQLite ledger with atomic caption claims and publication checkpoints

Verification

  • PYTHON_DOTENV_DISABLED=1 LEOMA_EVAL_TOKEN= OBJECT_STORAGE_BACKEND=r2 venv/bin/pytest -q: 646 passed
  • new corpus-v2 and CLI tests: 24 passed
  • uv lock --check: passed
  • uv build --wheel: passed
  • git diff --check: passed
  • pyflakes reports only five pre-existing warnings outside this change

Prepare deterministic five-second TI2V samples with content-addressed clips, lossless first frames, and a restart-safe SQLite ledger.

Add pinned eight-GPU Qwen captioning, human-QA publication gates, verified Hippius uploads, and sharded root-last manifests. Keep the active v1 evaluator and consensus configuration unchanged until evaluator-v2 support lands.
@vex0209-bt
vex0209-bt merged commit 5d0810c into main Jul 24, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant