Given a recipe and a dietary restriction (vegan, gluten-free, dairy-free, etc.), Robuchan produces a structured adaptation with ingredient substitutions, updated steps, flavor preservation notes, and a compliance self-check.

Built for the Mistral AI Worldwide Hackathon Tokyo (Feb 28 - Mar 1, 2026).

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Ministral-8B-Instruct-2410",
    device_map="auto",
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(base_model, "sumitdotml/robuchan")
tokenizer = AutoTokenizer.from_pretrained("sumitdotml/robuchan")

messages = [
    {
        "role": "system",
        "content": (
            "You are a culinary adaptation assistant. "
            "Priority: (1) strict dietary compliance, (2) preserve dish identity and flavor profile, "
            "(3) keep instructions practical and cookable. "
            "Never include forbidden ingredients or their derivatives (stocks, sauces, pastes, broths). "
            "If no exact compliant substitute exists, acknowledge the gap, choose the closest viable option, "
            "and state the trade-off. "
            "Output sections exactly: Substitution Plan, Adapted Ingredients, Adapted Steps, "
            "Flavor Preservation Notes, Constraint Check."
        ),
    },
    {
        "role": "user",
        "content": (
            "Recipe: Mapo Tofu\n"
            "Cuisine: Sichuan Chinese\n"
            "Ingredients: 400g firm tofu, 200g ground pork, 2 tbsp doubanjiang, "
            "1 tbsp oyster sauce, 3 cloves garlic, 1 inch ginger, 2 scallions, "
            "1 tbsp cornstarch, 2 tbsp neutral oil\n"
            "Steps: 1) Brown pork in oil until crispy. 2) Add minced garlic, ginger, "
            "and doubanjiang; stir-fry 30 seconds. 3) Add tofu cubes and 1 cup water; "
            "simmer 8 minutes. 4) Mix cornstarch slurry and stir in to thicken. "
            "5) Garnish with sliced scallions.\n"
            "Restrictions: vegetarian, shellfish-free\n"
            "Must Keep Flavor Notes: mala heat, savory umami, silky sauce"
        ),
    },
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
inputs = inputs.to(model.device)
outputs = model.generate(inputs, max_new_tokens=1024, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Output Format

The model produces five structured sections:

Section	Content
Substitution Plan	One row per banned ingredient: `original -> replacement (rationale)`
Adapted Ingredients	Full ingredient list with quantities — no placeholders
Adapted Steps	Complete numbered cooking steps reflecting all substitutions
Flavor Preservation Notes	3+ notes on how taste/texture/aroma are maintained
Constraint Check	Explicit checklist confirming all violations resolved

Training

Detail	Value
Base model	`mistralai/Ministral-8B-Instruct-2410`
Method	QLoRA SFT via TRL on HF Jobs (A10G)
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`
Training examples	1,090
Validation examples	122
Epochs completed	~0.95 (OOM at epoch boundary eval on A10G 24GB)
Final train loss	0.77

Dataset

Training data was synthetically generated from Food.com's 530K recipe corpus:

Filter source recipes that violate at least one supported dietary constraint
Generate structured adaptations using mistral-large-latest
Score each candidate with deterministic quality checks (constraint compliance, ingredient relevance, structural completeness)
Keep only passing candidates — single candidate per recipe, drop on fail

The dataset covers 10 dietary categories: vegan, vegetarian, dairy-free, gluten-free, nut-free, egg-free, shellfish-free, low-sodium, low-sugar, low-fat.

Three prompt templates (labeled-block, natural-request, goal-oriented) at a 50/30/20 split prevent format overfitting.

Dataset: sumitdotml/robuchan-data

Evaluation

Three-layer evaluation: format compliance (deterministic header parsing), dietary constraint compliance (regex against banned-ingredient lists), and LLM-as-judge via mistral-large-latest.

Metric	Baseline (`mistral-small-latest`, n=50)	Robuchan (n=3)	Delta
Format Compliance	14%	100%	+86pp
Constraint Compliance	0%	33%	+33pp
Judge Overall Score	9.20/10	—	—

Key findings:

The base model writes fluent recipe adaptations but fails at structured output (only 14% contain all 5 required sections) and completely fails dietary compliance (0% pass the banned-ingredient check).
Robuchan fixes structured output (100%) and begins enforcing dietary constraints (33%), though more training would likely improve compliance further.
The LLM judge overestimates compliance (9.88/10 for the base model despite 0% deterministic pass) — it measures attempt quality, not correctness.

W&B: sumit-ml/robuchan

Limitations

Small eval sample. Only 3 rows were evaluated on the fine-tuned model before the HF Space crashed. Results are directionally strong but not statistically robust.
Partial training. The adapter was saved from ~95% through epoch 1. More training would likely improve constraint compliance.
English only. Training data and evaluation are English-language recipes only.
Not safety-tested. This model is a hackathon prototype. Do not rely on it for medical dietary advice (severe allergies, celiac disease, etc.).

Authors

Framework Versions

PEFT: 0.18.1
TRL: 0.29.0
Transformers: 5.2.0
PyTorch: 2.6.0+cu124
Datasets: 4.6.1

Citation

bibtex

@misc{robuchan2026,
  title  = {Robuchan: Recipe Dietary Adaptation via Fine-Tuned Ministral-8B},
  author = {sumitdotml and Hiware, Kaustubh},
  year   = {2026},
  url    = {https://huggingface.co/sumitdotml/robuchan}
}

sumitdotml/robuchan

Robuchan