ADR-0433: CHUG Content Splits And HDR Audit¶
- Status: Accepted
- Date: 2026-05-15
- Deciders: Lusoris maintainers
- Tags: ai, hdr, chug, training, fork-local
Context¶
CHUG rows are bitrate-ladder variants of shared UGC-HDR source content. Splitting individual rows independently would allow the same source content to appear in both training and validation, inflating HDR experiments before any model decision is involved. The local CHUG pipeline also needs a pre-training HDR metadata check so malformed transfer / primaries signalling is caught before feature rows are used.
Decision¶
ai/scripts/chug_extract_features.py will assign deterministic train/validation/test partitions by hashing chug_content_name with seed chug-hdr-v1 into an 80/10/10 split. Every materialised feature row records split, chug_split_key, and chug_split_policy. The same script will optionally write a local ffprobe-backed HDR metadata audit JSON before feature extraction. train_konvid_mos_head.py will consume those explicit split labels when present, using train rows for fitting and val (or test when no val exists) as the held-out evaluation set. The FR-from-NR full-feature parquet extractor will accept an optional --metadata-jsonl sidecar so CHUG runs keep the same content and split metadata in parquet outputs. train_konvid_mos_head.py will also accept those FULL_FEATURES parquet files directly via --feature-parquet, mapping <feat>_mean aggregates back to the trainer's canonical feature columns.
Alternatives considered¶
| Option | Pros | Cons | Why not chosen |
|---|---|---|---|
| Row-level random split | Simple and balances row counts closely | Leaks bitrate-ladder variants of the same source content across validation | Rejected; leakage is worse than minor split imbalance |
| Manifest-order split | Deterministic without hashing | Depends on upstream CSV ordering and can cluster related content accidentally | Rejected; hash partitioning is stable and order-independent |
| Separate audit script | Keeps the materialiser narrower | Operators can forget the audit before training; duplicated JSONL loading / clip path handling | Rejected; the audit is a pre-training guard for the same local feature job |
Consequences¶
- Positive: CHUG HDR experiments get leakage-resistant splits and a repeatable metadata audit without waiting on any model work.
- Negative: Small local
--max-rowsruns may not contain all three splits because the split unit is content, not row. - Neutral / follow-ups: CHUG calibration jobs can now train from JSONL materialiser rows or FULL_FEATURES parquet rows; production HDR status still depends on the final teacher/model decision.