Fixing the Extra-Ingredient Problem in MiniChefGPT: Which Model Should We Use?

Description and use

MiniChefGPT produces recipes containing ingredients not supplied by the user. The solution is to replace it with SmolLM2‑360M‑Instruct, fine‑tune via LoRA, and use constrained decoding to reduce extra ingredients.

Problem

MiniChefGPT generates recipes that include ingredients not provided by the user, reducing recipe relevance.

Outcome

The plan is to replace the from‑scratch model with SmolLM2‑360M‑Instruct fine‑tuned via LoRA, add constrained decoding, and measure improvement via coverage and extra‑ingredient rate. Success will be confirmed when extra‑ingredient rate drops on held‑out lists.

Workflow

• 1. Convert recipe CSV into instruction examples: "Use only these ingredients: … Write a title and steps."
• 2. Fine‑tune SmolLM2‑360M‑Instruct with LoRA on the same training split.
• 3. Generate recipes using a banned‑word list derived from training food words minus allowed pantry staples.
• 4. Evaluate coverage and extra‑ingredient rate on held‑out lists and custom tests.

Evidence

We gave the app four ingredients: chicken, garlic, onion and tomato. It returned a recipe titled "chicken soup" whose steps use broccoli, rice, chicken broth and noodles.

Source

On four custom ingredient lists, the generated recipes mention 56% of the requested ingredients on average. On 12 held‑out ingredient lists the figure is 27%.

Source

I would replace our from‑scratch model with SmolLM2‑360M‑Instruct, fine‑tuned on our recipes with LoRA.

Source

In fp16, 360 million parameters take about 0.7 GB (parameters times 2 bytes), which fits easily on a free Colab T4 with 16 GB.

Source