tristanfaure.com /projects / recipe-llm-benchmark

Recipe LLM Benchmark

Qwen3-0.6B: Few-Shot vs RAG vs LoRA fine-tune vs RAG+FT, measured

2026 · ACADEMIC

Stack

Python · Qwen3 · LoRA (PEFT) · RAG · Hugging Face · Embeddings

Key figures

Strategies
4 compared
Corpus
30K recipes
Metrics
ROUGE · cosine · latency
Model
Qwen3-0.6B

A controlled benchmark of the four ways to specialize a small LLM: few-shot prompting, RAG over 30K recipes (embedding retrieval), LoRA fine-tuning, and RAG combined with fine-tuning.

Every strategy is scored on the same axes — ROUGE-1, semantic cosine similarity, and latency — so the trade-offs are numbers, not opinions.

The trained LoRA adapter ships with the repo; the same model family and adaptation stack later powered the AutoPilot AI chatbot on this site.

Links

All 33 projects · Open the interactive portfolio · About Tristan