Research 0108 — KonViD-150k Split Score Layout¶
Research-0108¶
Date: 2026-05-14
Question¶
The existing KonViD-150k adapter expected one manifest.csv with URLs, MOS, optional standard deviation, and rating count. The local corpus drop uses k150ka_scores.csv / k150kb_scores.csv plus k150ka_extracted/ / k150kb_extracted/, so the adapter could not consume the corpus already staged under .workingdir2/konvid-150k/.
Finding¶
The split score layout is a real upstream distribution shape, not a test fixture. k150ka_scores.csv carries video_name,video_score; k150kb_scores.csv carries video_name,mos,video_score. The staged directories hold the matching MP4s, so no URL reconstruction is needed.
The MOS-corpus JSONL schema should not widen for this layout. The shared trainers depend on the existing row shape, and split score CSVs do not provide per-row standard deviation or rating counts. The correct adapter behavior is to preserve mos, fill mos_std_dev = 0.0, n_ratings = 0, and probe geometry from the local MP4s exactly as the manifest path does.
Alternatives Considered¶
| Option | Result | Reason |
|---|---|---|
Require operators to synthesize manifest.csv manually | Rejected | Leaves the in-tree adapter unable to consume the common score-drop layout and repeats fragile local conversion logic. |
Add separate --split-score-layout CLI mode | Rejected | The default directory has an unambiguous discovery order: use explicit/real manifest.csv first, otherwise discover split score CSVs. |
Widen JSONL with split / score-source columns | Rejected | Downstream MOS-corpus consumers rely on the shared schema. Split identity is useful for diagnostics but not part of the trainer contract. |
Auto-discover split score CSVs when manifest.csv is absent | Chosen | Consumes the staged corpus without schema drift, while explicit --manifest-csv remains strict so typoed paths fail loudly. |