Skip to content

Research 0648: CHUG HDR MOS Trainer Surface

Question

Can CHUG HDR subjective-MOS training reuse the existing small MOS-head training loop without exposing CHUG through a KonViD-named command?

Findings

  • CHUG feature rows already carry trainer-ready fields: mos mapped to [1, 5], mos_raw_0_100, canonical feature columns such as adm2 and motion2, and content-level split labels.
  • The existing MOS trainer's loader can consume JSONL rows and preserves explicit split labels, but the public CLI had only KonViD-named JSONL inputs.
  • The same model architecture is usable for a local CHUG HDR MOS experiment, but the model identity must not be konvid_mos_head_v1; CHUG is HDR MOS, while the current Netflix teacher is SDR/8-bit calibrated.

Result

The lowest-risk implementation is a CHUG-named wrapper, ai/scripts/train_chug_hdr_mos_head.py, over the shared MOS-head loop. It gives operators a truthful command and defaults while preserving the committed KonViD model's existing script and provenance.

Verification

Regression coverage adds a CHUG-wrapper smoke path that trains from synthetic CHUG-style feature JSONL, writes a local ONNX/manifest pair, and asserts the manifest id is chug_hdr_mos_head_v1.