Daniel-Guckert/brie-3b

Brie 3B

A cheese-specialist chat model, named Brie: Qwen2.5-3B-Instruct QLoRA-fine-tuned on ~14.5k generated examples, paired with a small local cheese database (1,200+ cheeses) for factual grounding.

The model alone will mostly say "not in my database" for anything specific — the database and retrieval script are required for real answers. This repo ships all three.

Results

• Train loss: 0.518 | Eval loss: 0.152 | Eval token accuracy: 95.4% (2 epochs, held-out eval split)
• Base: Qwen2.5-3B-Instruct, QLoRA rank 32, all attention + MLP layers
• Trained on an RTX 2060 (6GB), ~4h52m

Files

File	Size	What it's for
cheese-3b-q4_k_m.gguf	~1.9GB	Recommended — good quality/size for CPU or small GPU
cheese-3b-q5_k_m.gguf	~2.3GB	Higher quality, still small
cheese-3b-f16.gguf	6.2GB	Full precision, for further quantizing or fine-tuning
cheese.db	~2MB	SQLite database of 1,200+ cheeses (country, milk, texture, aging, etc.)
cheese_rag.py	—	Retrieval layer: lookups, filtered lists ("sheep milk cheeses from France"), counts
cheese_chat.py	—	CLI chat client wiring the database into the model
Modelfile	—	Ollama modelfile

Quick start (Ollama)

ollama create brie -f Modelfile
python cheese_chat.py --model brie

Quick start (any OpenAI-compatible server, e.g. llama.cpp's llama-server)

llama-server -m cheese-3b-q4_k_m.gguf --port 8080
python cheese_chat.py --openai http://localhost:8080/v1 --model brie

cheese_chat.py is stdlib-only Python — no installs beyond Python itself.

What it's good at

• Cheese facts covered by the database: origin, milk type, texture, aging, fat, certification
• Listing / counting cheeses by filter ("how many French cheeses do you have?", "list 20 goat's milk cheeses")
• Staying on-topic and refusing non-cheese requests
• Honestly saying when the database has no entry or no value for a field, instead of guessing

What it's not

• A source for cheeses outside its ~1,200-entry database
• A substitute for professional food-safety, allergen, or nutrition advice

License

• Model weights: derived from Qwen2.5-3B-Instruct under the Qwen Research License. Non-commercial research/personal use. Built with Qwen.
• Cheese database: derived from Wikipedia content, CC BY-SA 4.0. Attribution: Wikipedia contributors.
• Code (cheese_rag.py, cheese_chat.py): MIT — do whatever you want with it.

This is a hobby project, not an official or commercial release.