One clip, four voices. A Gemma 3 4B that learned tone through judge-preference training on AMD MI300X — riding inside a CPU-only container that survives anything.
These captions are unedited output from the exact Docker image submitted for evaluation, run against the organizers' three example clips.
| arm | formal | sarcastic | humorous_tech | humorous_non_tech | ALL |
|---|---|---|---|---|---|
| Kimi K2.6 prompted | 8.92 | 7.72 | 8.23 | 7.87 | 8.18 |
| Gemma 3 4B base | 7.88 | 6.78 | 8.18 | 8.02 | 7.72 |
| Gemma tuned v1 | 7.90 | 7.32 | 7.82 | 7.97 | 7.75 |
| Gemma tuned v2 | 8.40 | 7.13 | 7.93 | 7.83 | 7.83 |
| v2 vs base | +0.52 | +0.35 | −0.25 | −0.19 | +0.11 |
30 held-out scenes × 4 styles, judged by gpt-oss-120b — an independent model family, so nobody grades their own homework. Two targeted DPO rounds: round 1 lifted sarcastic +0.54 over base; round 2 lifted formal +0.50 — and brought the 4B to frontier parity on factual accuracy (7.62 vs Kimi's 7.62). The entire residual gap is tone (8.03 vs 8.74): what resists distillation into 4B parameters is wit. So best-of-N ships as default and the tuned Gemma rides in-container as the zero-dependency fallback voice.
ffmpeg single-pass frames + Whisper transcript
Kimi K2.6 vision → dense factual description
Gemma 3 4B — DPO-tuned, quantized, on CPU inside the container
LLM judge scores K candidates → best per style
Pre-written results, atomic upgrades, 3-tier fallbacks, 10-min budget