v0.3 Quality-Hardened Release Notes
- Seven configurations, 1,000 rows each (950 train / 50 eval).
- Schema:
streaming_omni_v0.3.1. - Timeline chunk IDs, Proactive evidence resolution, and OCR grounding/target resolution cover all applicable rows.
- Qwen2.5-VL, WhisperX, joint-modality, reconstruction, SFT, construction, publication, and remote reports are the authoritative source for final measured values.
- Earlier fallback-parsed VLM captions are not accepted as fully structured output; affected sources are re-inferred and validated.
- OCR is generator ground truth, and WhisperX timing is segment-level.
Measured hardening results:
- Qwen2.5-VL-7B: 350/350 real re-invocations; 0 direct JSON, 350 safely repaired JSON, 0 validated partial, 0 invalid/empty; every accepted result passed the required structured-field checks.
- CLAP: 350 existing real audio-tagging outputs retained.
- WhisperX: 350/350 real processing records audited; 339/339 speech-present clips have non-empty transcription, 11 explicit no-speech outcomes, 0 wrapper/timestamp errors.
- Joint modality: 1,000 rows structurally checked and 100 rows agent-reviewed; 43.0% meet the conservative genuinely-joint rule and audited false-joint rate is 0%.
- ProactiveAVInteraction: 1,000 rows, 950/50, 350 unique sources.
- OCRScreenGrounding: 1,000 rows, 950/50, 315 unique sources; generator-ground-truth grounding and target resolution are 100%.