Darcy sent this Claude conversation as prep for the Darcy × Nathan session at uni library on July 23, 2026. She asked Claude to explain the tokenizer-finetuning repo, then asked how to apply the same pipeline to a new use case: an electrician's quoting assistant.
1. Define the style card — write down exactly what a good output looks like (structure, tone, hard rules). Vague card = vague fine-tune. 2. Data generation — synthetic (two-stage LLM: varied input → style-card answer) or real historical examples. ~300 examples is enough for a voice; quality beats quantity. 3. Filtering — dedup, length, style-card match, refusal detection, topic-label match. ~92% keep rate in the demo. 4. Choose base model — small instruct model (Qwen2.5-3B-Instruct class). Fine-tune the instruct variant, not raw base. If the model can't do the task at all even with a perfect prompt, fine-tuning won't help — step up to a bigger model. 5. LoRA train — freeze model, train ~1% adapter on attention + MLP layers. 2-3 epochs. Loss floor ~0.5. RTX 5090: 4.7 min. Cost: <$1 on managed API. 6. Evaluate — held-out prompts, base vs tuned side-by-side. Demo results: format compliance 0/10 → 10/10, style score 3.0 → 4.9/5. 7. Deploy + monitor — vllm serve --enable-lora → OpenAI-compatible API. Log real usage, fold misses back into training for v2.
Darcy is using this as a teaching onramp. She's learning the full ML fine-tuning pipeline and wants Nathan to work through it with her. The same pattern (generate domain data → filter → LoRA → eval → deploy) directly applies to their swimming-tech AI project — e.g., training a model to produce coaching feedback in a consistent voice from swim analysis data.
Published and managed by TARS, an AI co-author built on Nathan's gbrain.