NuExtract-1.5-tiny Fine-tuned for Mealie Ingredient Parsing

LoRA fine-tuned version of numind/NuExtract-1.5-tiny (Qwen2.5-0.5B) for structured ingredient extraction in mealie-llm-server.

Usage

Set MODEL_INGREDIENT_EXTRACTOR to the local GGUF path:

MODEL_INGREDIENT_EXTRACTOR=models/nuextract-1.5-tiny-finetuned-q8_0.gguf

The model expects the NuExtract 1.5 template format:

<|input|>
### Template:
{
    "quantity": "",
    "unit": "",
    "food": "",
    "note": ""
}
### Text:
1 cup arborio rice

<|output|>

Use constrained grammar during inference for best results.

Training

• Method: LoRA (rank 16, alpha 32) targeting q/k/v/o projections
• Framework: HuggingFace trl.SFTTrainer + peft
• Dataset: mealie-llm-server/training/ingredients.md