LVBench 4,000-candidate Video Crop v8 training archive
Private archive of the completed first LVBench production batch: v8 prompts, per-sample streaming pipeline, and an independent raw Gemini GT-matched export.
Download LVBench_4000_v8_training_bundle_v1.zip and its .zip.sha256 file, verify the checksum, and extract the ZIP without changing its directory structure.
The bundle includes:
- 6,369 baseline training rows: 1,752 complete trajectories, 1,901 partial prefixes, and 2,716 salvage prefixes.
- 8,579 weighted rows.
- 2,873 original Gemini GT-matched trajectories exported without Seed post-filtering.
- All nine product variants in JSONL and Parquet, with aligned metadata and sampling weights.
- 164,507 images, including all media referenced by the products and the included trajectory histories.
- Tool definitions, a portable JSONL reader, usage instructions, and a per-file SHA-256 manifest.
Media paths are relative to the extracted bundle root. SFT data loading requires no access to the original server or generation APIs. Model weights, tokenizer, and training framework are supplied by the training environment.
Baseline, weighted, and category-specific products overlap; choose the intended variant rather than concatenating all files. Raw GT products may overlap the reviewed products as well.
No API keys, authentication certificates, or raw API request/response files are included. Source questions and media retain their upstream terms.
sha256sum -c LVBench_4000_v8_training_bundle_v1.zip.sha256
unzip LVBench_4000_v8_training_bundle_v1.zip
python LVBench_4000_v8_training_bundle_v1/training_data.py LVBench_4000_v8_training_bundle_v1
- Downloads last month
- 29