BRIE

(a small-data fine-tuning study)

The same 1,213 hand-written examples, fine-tuned into four small models (0.5B–7B) holding one continental-philosophy register. Blind frontier-model judges prefer the fine-tunes to their base models, with no measurable out-of-domain regression. Total compute, at most, three dollars.

The 0.5B model was small enough to train on a laptop.

medium: LoRA r=16, four base models, ~$3 of GPU
year: 2025–2026
venue: Zenodo, DOI 10.5281/zenodo.17657737
duration: two epochs
tags: fine-tune · method · autobiography

→ paper (DOI 10.5281/zenodo.17657737)
→ code & evaluation framework
→ model · Brie Qwen 2.5 7B
→ model · Brie Qwen 2.5 3B
→ model · Brie Llama 3.2 3B
→ model · Brie Qwen 2.5 0.5B