kerala-recipes-qwen-3b

🥥 Kerala Recipes — Qwen 3B Fine-Tuned Model

A fine-tuned Qwen2.5-3B-Instruct model that knows 89 authentic Kerala recipes by heart.

Ask it any Kerala dish — it gives you the exact recipe with scaled ingredient amounts, step-by-step method, and meal pairings. No long intros, no fluff. Just the recipe, straight and simple.

What This Model Does

• 🍛 Generates complete Kerala recipes with exact gram/ml quantities
• 👥 Scales ingredients for any number of people (1 to 30)
• 🌶️ Covers curries, breakfast items, rice dishes, snacks, desserts, and drinks
• 🍽️ Suggests what dishes go well together (meal pairings)
• 🗣️ Understands both English and Malayalam dish names

Try It Out

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Fathi7ma/kerala-recipes-qwen-3b")
tokenizer = AutoTokenizer.from_pretrained("Fathi7ma/kerala-recipes-qwen-3b")

messages = [
    {"role": "system", "content": "You are a concise Kerala cooking assistant. Give the recipe directly with exact ingredient amounts and simple steps. No intro talk or fluff."},
    {"role": "user", "content": "How do I cook Nadan Meen Curry for 4 people?"},
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_ids=inputs, max_new_tokens=512, temperature=0.3, repetition_penalty=1.15)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Example Output

Prompt: "How do I cook Nadan Meen Curry for 4 people?"

Model Response:

Nadan Meen Curry (നാടൻ മീൻ കറി) Non-Vegetarian · Spice 3/5 · Prep 10m · Cook 20m

Serves 4

Ingredients:

• 600.0 g Mackerel / sardine, cleaned
• 80.0 g Shallots, sliced
• 12.0 num Green chillies
• 12.0 g Ginger-garlic paste
• 16.0 g Chilli powder
• 4.0 g Turmeric powder
• 40.0 g Coconut oil
• 700.0 g Water

Method:

1. Marinate fish in turmeric, salt and pepper; rest 20 minutes.
2. Heat oil, splutter curry leaves and fry shallots until golden.
3. Add ginger-garlic and chillies; saute till they turn soft.
4. Add chilli, turmeric and pepper; cook until aromatic.
5. Add water and bring to a boil.
6. Add the fish; simmer gently for 8-12 minutes until cooked through.
7. Garnish with fresh curry leaves and serve with rice.

Goes well with: Steamed Rice, Kappa, Thoran

What's Inside

The Dataset

This model was trained on Fathi7ma/kerala-recipes, a hand-curated collection of 89 traditional Kerala dishes. Each dish includes:

• Base quantities per 1 serving with scaling formulas (linear, fractional, fixed)
• Dual units (grams + kitchen measures like cups, tablespoons)
• Step-by-step cooking instructions
• Meal pairing suggestions (what goes well with what)
• Malayalam dish names

The recipes come from home kitchens in Kerala — the kind of measurements and methods passed down through families.

Dishes Covered

Curries, stews, biryanis, breakfast items (puttu, appam, dosa, idli), rice dishes, snacks (vada, bajji, unniyappam), side dishes (thoran, mezhukkupuratti, pachadi), desserts (payasam, halwa), and drinks (sambharam, sulaimani).

Try the Live Demo

👉 Kerala Recipes Demo Space — browse all 89 recipes with ingredient scaling

Training Details

• Hardware: Google Colab T4 GPU (free tier)
• Training Time: ~5 minutes
• Epochs: 3
• LoRA Rank: 16
• Learning Rate: 2e-4
• Batch Size: 2 (with 4x gradient accumulation)

Limitations

• Trained on 89 Kerala dishes — may not know recipes outside this collection
• Ingredient scaling uses simplified formulas; some items like salt may need personal adjustment
• Best results when using the system prompt shown in the example above

Author

Made by Fathi7ma — a Kerala food enthusiast who wanted to preserve authentic home-cooking recipes in a format that AI can understand and share.