food-assistant

Food Assistant – Falcon3-7B-Instruct (Fine-tuned with QLoRA)

Model Description

This model is a fine-tuned version of Falcon3-7B-Instruct designed to extract nutritional components from a given food name. It returns a structured list of nutrients (proteins, fats, vitamins, minerals, etc.) based on a curated food composition dataset.

The fine-tuning was performed using QLoRA (Quantized Low‑Rank Adaptation) to achieve efficient training with limited GPU memory.

---

Model Details

Model Description

• Developed by: Maigamhdsidy
• Funded by: Academic project (I3AFD Course)
• Shared by: Maigamhdsidy
• Model type: Causal Language Model fine‑tuned for structured information extraction
• Language(s) (NLP): English
• License: Apache‑2.0
• Finetuned from model: tiiuae/Falcon3-7B-Instruct

Model Sources

• Repository: https://huggingface.co/maigamhdsidy/food-assistant
• Demo: You can test it via the inference widget on this page.

---

Uses

Direct Use

You can use this model to:

• Automatically extract nutrient lists from food names.
• Enrich food databases or recipe applications.
• Assist in dietary planning and nutritional analysis.

Downstream Use

The model can be integrated into chatbots, mobile apps, or web services that need quick nutritional information.

Out‑of‑Scope Use

• The model is not a general‑purpose conversational AI.
• It should not be used for medical diagnosis or clinical nutrition advice.
• It may hallucinate values for foods not present in the training set.

---

Bias, Risks, and Limitations

• Data bias: The training data comes from a specific dataset that may not cover all foods or cuisines.
• Hallucination: For unknown foods, the model may generate plausible but incorrect nutrient values.
• Units: The model outputs values with units (g, mg, kcal, etc.) but does not guarantee precision.
• Recommendations: Always verify generated nutritional data with trusted sources before using in critical applications.

---

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Base model + adapter
base_model_name = "tiiuae/Falcon3-7B-Instruct"
adapter_path = "maigamhdsidy/food-assistant"

tokenizer = AutoTokenizer.from_pretrained(adapter_path)

model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",
    trust_remote_code=True
)
model = PeftModel.from_pretrained(model, adapter_path)

# Example prompt
prompt = """Given the following food name, return only its nutrient components.

Food name: Banana

Required output:
- Return only the components list."""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))