A QLoRA fine-tune of Qwen/Qwen2.5-7B-Instruct for opinionated research triage: it reads an item (HF model/dataset/paper) and emits a structured verdict — read-now | skim | skip | archive + novelty score + one-line reason + tags.

Trained on a Sonnet-4.6-distilled dataset (true distillation, not self-distillation) via PEFT+TRL (bf16 LoRA). Part of a 4-model bake-off: every fine-tune was judged blind against its own base by a 3-lens panel (decisiveness / specificity / calibration).

Training data: hqt2yotoz/verdict-engine-triage-v2 — 468 train + 52 val examples, Sonnet-4.6-distilled (true distillation).

Fine-tune 10 / base 0 over 10 held-out firehose items — beats base. Full leaderboard:

The base over-praises (read-now + inflated novelty); the fine-tune discriminates (skim + honest novelty). The calibration panel rewards the fine-tune.

Built for the HuggingFace Build Small Hackathon. Repo: github.com/setuc/verdict-engine