AMD Developer Hackathon: ACT II — Track 2

StyleForge

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.

DPO fine-tuned Gemma 3 4B Trained on AMD Instinct MI300X Kimi K2.6 vision via Fireworks Best-of-N judge reranking

The example clips, captioned by the shipped container

loading…

These captions are unedited output from the exact Docker image submitted for evaluation, run against the organizers' three example clips.

Measured honestly

armformalsarcastichumorous_techhumorous_non_techALL
Kimi K2.6 prompted8.927.728.237.878.18
Gemma 3 4B base7.886.788.188.027.72
Gemma tuned v17.907.327.827.977.75
Gemma tuned v28.407.137.937.837.83
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.

Architecture: eyes, voice, taste

Ingest

ffmpeg single-pass frames + Whisper transcript

Eyes

Kimi K2.6 vision → dense factual description

Voice

Gemma 3 4B — DPO-tuned, quantized, on CPU inside the container

Taste

LLM judge scores K candidates → best per style

Armor

Pre-written results, atomic upgrades, 3-tier fallbacks, 10-min budget

Sankar Subbayya × Claude (Anthropic) — July 2026 github.com/SankarSubbayya/styleforge