CulinaAI - Qwen2.5-7B Recipe Generation (GGUF, Q4_K_M)

Model Details

• Base model: Qwen/Qwen2.5-7B-Instruct
• Fine-tuning method: LoRA (QLoRA-style, 4-bit base model)
• Training framework: Unsloth Studio
• Format: GGUF, quantized to Q4_K_M (~4.68 GB)
• Lifetime Downloads: 1,000+ (Note: The standard Hugging Face header metric only logs the trailing 30 days)

Training Configuration

• Smoothed training loss: 0.941
• Best eval loss: 0.719

Usage

• llama.cpp
• Ollama
• LM Studio
• Unsloth Studio itself

Ollama

Example Prompt & Output

• 1 cup chilies, ground
• 2 tablespoons oil
• 3/4 cup all-purpose flour
• 680.4 g paneer cubes
• salt to taste
• 1 (250 ml) can tomato puree
• 3/4 teaspoon chili powder
• 1/2 teaspoon turmeric

1. Heat oil in a wok and fry the chilli paste till brown.
2. Drain well on absorbent paper.
3. Make a smooth batter with flour, chilli paste, salt, tomato puree, turmeric and chilli powder.
4. Heat 6 tbsp of ghee (or butter) in a large pan.
5. Add the paneer cubes and fry till golden brown.
6. Remove from heat.
7. Beat the batter well and deep-fry it until crisp.

Limitations

• Generated recipes may contain inaccurate quantities, cooking times, or steps. Always use judgment and review before cooking.
• The model may occasionally hallucinate ingredients or techniques that aren't standard.
• Trained with a 1024-token context window, so very long recipes or multi-recipe requests may be truncated.

Training Data

• title
• meal_type, primary_classification, secondary_classification, context_dependent_classification — recipe categorization
• ingredients, NER — ingredient lists (raw and NER-extracted entities)
• instructions — step-by-step cooking instructions

Note: This repository contains only the fine-tuned model weights, not the training dataset itself. Recipe1M+ is distributed under MIT's own terms of access — refer to the official Recipe1M+ site for dataset licensing if you plan to redistribute the data.