The code package distributes construction and validation code. Training records, source media, benchmark exports, and calibration artifacts are external inputs.
The trainer consumes a sealed semantic release. Its validation checks the manifest, source inventory, user/assistant message schema, media identity, and permitted targets. See:
viss_core/data_v0/schema.pyviss_core/data_v0/semantic_release.pyviss_core/data_v0/trainer_adapter.py
The source-aligned builder combines already prepared source releases:
python -m data_pipeline.temporal_scope.build_source_aligned_sft_release \
--source data/prepared_source_a \
--source data/prepared_source_b \
--output data/viss_release \
--release-id my_viss_release \
--copy-mode copyprepared_source_a and prepared_source_b are placeholders for valid
source releases, not raw-video directories. This generic example does not
reconstruct the paper's accepted training mixture by itself.
The paper describes the training mixture. Company training records and their manifests are not distributed. The training guide assumes users have prepared their own valid source releases, media, example weights, and ordering.
Set VISS_UPSTREAM_MANIFEST_SHA256 to the trusted upstream identity for your
training inputs before starting the builder or trainer. The validator retains
the identity check, but the code package embeds no company-specific digest.
Private legacy source adapters are excluded; supply the standard semantic
release format instead.
For each benchmark, prepare:
- Official annotations and authorized video files.
- A query export using the fields read by
viss_core/benchmarks.py. - The benchmark's official scoring installation.
- Model and processor files compatible with the locked evaluation protocol.
configs/evaluation/recent4.example.json describes the public sampling,
query-count, prompt, and processor conventions. It contains no run dates,
company file identities, or query records. Follow the evaluation guide to
create a local configuration bound to your own query exports and processor.
The generated configuration is excluded from publication.
The TVS policy consumes acquisition/commit calibration artifacts and aligned view predictions from one model. Calibrate on source-disjoint data, then freeze the thresholds for benchmark evaluation. The archive includes the selection and analysis implementation, but not the paper's calibration records or precomputed predictions.