---
language:
- en
license: apache-2.0
base_model: Qwen/Qwen3-0.6B
tags:
- recipe-generation
- food
- cooking
- fine-tuned
- qwen3
- unsloth
datasets:
- recipe_nlg
pipeline_tag: text-generation
---

# Qwen3-0.6B Recipe Chef

A fine-tuned version of [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) trained on 60,000 recipes from the RecipeNLG dataset.
Give it a list of ingredients and it generates a complete recipe with title, quantities, and step-by-step directions.

## Model Details

| Property        | Value                        |
|----------------|------------------------------|
| Base model      | Qwen/Qwen3-0.6B              |
| Training data   | RecipeNLG (70k samples)      |
| Fine-tune method| LoRA (r=64, alpha=128)        |
| Epochs          | 2                            |
| Training loss   | 0.86                        |
| Framework       | Unsloth + TRL                |

## How to Use

### Option 1 — With Unsloth (recommended, faster)

```python
from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "Aniq-63/qwen3-0.6B-recipe-finetuned",
    max_seq_length = 1024,
    load_in_4bit   = True,
)

FastLanguageModel.for_inference(model)

@torch.inference_mode()
def generate_recipe(ingredients: str) -> str:
    messages = [
        {
            "role": "system",
            "content": (
                "You are a professional chef assistant. "
                "When given a list of ingredients, generate a complete recipe with "
                "a title, structured ingredient list with quantities, and clear "
                "step-by-step directions."
            )
        },
        {
            "role": "user",
            "content": ingredients
        }
    ]
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False,
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens = 400,
        temperature    = 0.7,
        top_p          = 0.9,
        do_sample      = True,
        use_cache      = False,
    )
    new_tokens = outputs[0][inputs["input_ids"].shape[1]:]
    return tokenizer.decode(new_tokens, skip_special_tokens=True)

print(generate_recipe("chicken, garlic, onion, olive oil, tomato"))
```

### Option 2 — With standard Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Aniq-63/qwen3-0.6B-recipe-finetuned",
    torch_dtype    = torch.float16,
    device_map     = "auto",
)
tokenizer = AutoTokenizer.from_pretrained("Aniq-63/qwen3-recipe-chef")

messages = [
    {
        "role": "system",
        "content": (
            "You are a professional chef assistant. "
            "When given a list of ingredients, generate a complete recipe with "
            "a title, structured ingredient list with quantities, and clear "
            "step-by-step directions."
        )
    },
    {
        "role": "user",
        "content": "chicken, garlic, onion, olive oil, tomato"
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens = 400,
    temperature    = 0.7,
    top_p          = 0.9,
    do_sample      = True,
)

new_tokens = outputs[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))
```

## Training Details

- **Dataset:** [RecipeNLG](https://www.kaggle.com/datasets/paultimothymooney/recipenlg)
- **Fine-tune method:** LoRA (Unsloth)
- **Epochs:** 2