👨‍🍳 Chef-Mistral-7B: Constraint-Aware Recipe Generator

Chef-Mistral-7B is a fine-tuned language model designed to generate cooking recipes based on strict user constraints. Unlike general-purpose LLMs, this model is optimized to minimize "hallucinations" (inventing ingredients the user doesn't have) and strictly adhere to dietary restrictions (Keto, Vegan, Gluten-Free).

It was trained using QLoRA (Quantized Low-Rank Adaptation) on a consumer GPU (RTX 4070 Super) using the unsloth library for 2x faster training.

🚀 Capabilities

• Pantry-Based Cooking: Generates recipes using only the ingredients provided by the user.
• Dietary Adherence: Respects hard constraints like "No Dairy," "Keto," or "Gluten-Free."
• Context Awareness: Can distinguish between breakfast, dinner, and dessert contexts based on ingredients.

🏗️ Agent Architecture Recommendation

This model was specifically trained to serve as the reasoning engine (Brain) in a modular agentic system. For optimal performance, we recommend the following architecture:

• Front-End Router ("The Waiter"): Use a lightweight model (e.g., Phi-3, Qwen-1.5B) to handle user chit-chat and extract structured constraints (ingredients, diet) into JSON.
• The Brain (Chef-Mistral): Use this model only to generate the final recipe based on the structured input. It performs best when stripped of conversational filler.
• RAG Layer: Augment the prompt with retrieved safety facts (e.g., cooking temps, allergen data) to minimize hallucinations before generation.

💻 How to Use

You can load this model directly in Python using the unsloth or peft library.

Installation

pip install unsloth torch transformers

from unsloth import FastLanguageModel

# 1. Load the Model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "YourUsername/Chef-Mistral-7B", # REPLACEME WITH YOUR USERNAME
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# 2. Define the Prompt Template
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
"""

# 3. Run Inference
inputs = tokenizer(
    [
        alpaca_prompt.format(
            "I have chicken breast, spinach, and heavy cream. I want a keto-friendly dinner.", # Instruction
            "", # Input
            "", # Output - leave blank for generation
        )
    ], return_tensors = "pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens = 512, use_cache = True)
print(tokenizer.batch_decode(outputs)[0])

📊 Training Details

• Base Model: Mistral-7B-v0.3 (4-bit quantization)
• Method: QLoRA (LoRA Rank = 16, Alpha = 32)
• Dataset Size: ~25,000 instruction-response pairs (Recipe/Culinary dataset).
• Hardware: Trained on a single NVIDIA RTX 4070 Super (12GB VRAM).
• Training Time: ~3.5 hours (1545 Steps).
• Final Loss: ~0.65-0.7

⚠️ Limitations & Bias

• Hallucination: While significantly reduced, the model may occasionally hallucinate common pantry staples (e.g., assuming you have salt, pepper, or oil even if not listed).
• Adversarial Prompts: The model generally refuses illogical requests (e.g., "Chicken Dessert") but may exhibit repetition loops if forced into contradictory constraints.
• Medical Advice: This model is for culinary inspiration only. Do not rely on it for severe medical allergies.

👨‍💻 Author

Developed by [Hazar Utku Sozer] as part of the Deep Learning Engineering curriculum.