A lightweight LoRA adapter (≈ 200 MB) that teaches SmolLM3-3B to generate detailed, step-by-step Indian recipes given a dish name and ingredient list. Trained on the open-source EmTpro01/indian-recipe-cleaned corpus (6 871 recipes).

Model Details


Developer	Susant-Achary
Base model	`HuggingFaceTB/SmolLM3-3B`
Adapter type	LoRA (`r=16`, `α=32`, dropout 0.05)
Quantisation	4-bit NF4, bfloat16 compute (BitsAndBytes)
Languages	English (culinary domain)
License	Apache-2.0 (inherits base-model license)
Finetuning data	6 871 Indian recipes (CC-BY-SA-4.0)
Hardware	1 × A100-40 GB

Model Sources

Weights & tokenizer: this repository
Dataset: see link above

Uses

Direct Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
base  = "HuggingFaceTB/SmolLM3-3B"
lora  = "Susant-Achary/smollm3-indian-recipes"

tok   = AutoTokenizer.from_pretrained(lora)
model = AutoModelForCausalLM.from_pretrained(
            lora,
            load_in_4bit=True,
            device_map="auto",
            torch_dtype="bfloat16")

prompt = "Give me a detailed, step-by-step recipe for Paneer Butter Masala using these ingredients: paneer, tomato, butter, cream, garam masala."
print(tok.decode(model.generate(**tok(prompt, return_tensors="pt").to(model.device),
                                max_new_tokens=256)[0], skip_special_tokens=True))

Ask in Spanish , it still responds

#-----------------------------------------------------------------------------

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# 1.  Load the base + LoRA adapter (4-bit)
lora  = "Susant-Achary/smollm3-indian-recipes 
tok  = AutoTokenizer.from_pretrained(lora)
model = AutoModelForCausalLM.from_pretrained(
            repo,
            device_map="auto",
            load_in_4bit=True,
            torch_dtype=torch.bfloat16)

# 2.  Ask for a Spanish recipe
system  = "Eres un chef experto en cocina india. Responde siempre en español."
usuario = ("Dame una receta detallada, paso a paso, para hacer 'Chole Bhature' "
           "utilizando los siguientes ingredientes: garbanzos, cebolla, tomate, "
           "masala de garbanzos, harina de trigo, yogur, aceite.")
chat = tok.apply_chat_template(
        [{"role":"system", "content":system},
         {"role":"user",   "content":usuario}],
        tokenize=False, add_generation_prompt=True)

inputs = tok(chat, return_tensors="pt").to(model.device)
out_ids = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9)
print(tok.decode(out_ids[0][inputs.input_ids.shape[1]:],
                 skip_special_tokens=True))

Citation if you like it

bibtex

@misc{smollm3_indian_recipes_2025,
  title   = {SmolLM3-3B: Indian-Recipe LoRA},
  author  = {Susant-Achary},
  year    = {2025},
  howpublished = {HuggingFace Hub},
  url     = {https://huggingface.co/<susant-achary>/smollm3-indian-recipes}
}