minicpm5-1b-indian-fmcg-normalizer

MiniCPM5-1B — Indian FMCG Product Name Normalizer

Fine-tuned openbmb/MiniCPM5-1B to normalize inconsistent Indian FMCG product names from distributor invoices into canonical SKU names. Part of the Kirana Detective project — an AI system for small Indian grocery stores to reconcile purchase invoices against a master product catalog.

Model Description

Indian kirana (grocery) distributors write product names in highly inconsistent shorthand on invoices: abbreviations, typos, mixed case, regional variants, and truncated descriptions. This model maps any such raw string to a clean, canonical product name that can be matched against a standardized catalog.

Example inputs → outputs:

How to Use

With llama.cpp

llama-cli --model MiniCPM5-1B.Q4_K_M.gguf \
  -p "<|im_start|>system\nYou are an Indian FMCG product name normalizer. Given a raw product name from a distributor invoice, return ONLY the canonical product name. No explanation, no punctuation — just the canonical name.<|im_end|>\n<|im_start|>user\nInvoice product name: \"MAGGI NDL 70GM\"<|im_end|>\n<|im_start|>assistant\n"

With Python (transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("build-small-hackathon/minicpm5-1b-indian-fmcg-normalizer")
model = AutoModelForCausalLM.from_pretrained("build-small-hackathon/minicpm5-1b-indian-fmcg-normalizer")

messages = [
    {"role": "system", "content": "You are an Indian FMCG product name normalizer. Given a raw product name from a distributor invoice, return ONLY the canonical product name. No explanation, no punctuation — just the canonical name."},
    {"role": "user", "content": 'Invoice product name: "MAGGI NDL 70GM"'},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=32)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# → Nestle Maggi Masala Noodles 70g

Training Details

Dataset

• Size: 2,000 synthetic (raw, canonical) pairs covering ~200 Indian FMCG SKUs
• Train / Eval split: 1,800 / 200 (90 / 10)
• Generation method: Rule-based augmentation of a hand-curated catalog — abbreviation expansion, typo injection, random truncation, and regional shorthand patterns
• Format: Chat template (system + user + assistant turns)

Hyperparameters

Training Infrastructure

Quantization

The published file is MiniCPM5-1B.Q4_K_M.gguf (~688 MB), quantized from BF16 using llama.cpp Q4_K_M.

Limitations

• Trained on synthetic data only — real-world invoice noise may differ from the patterns simulated during generation
• Covers ~200 Indian FMCG SKUs; performance on out-of-catalog or regional/private-label products is untested
• Output is a direct string match; post-processing fuzzy matching against a catalog index is recommended in production

Citation

@misc{kirana-detective-2026,
  title  = {Kirana Detective: AI Invoice Reconciliation for Indian Grocery Stores},
  author = {Syed Naazim hussain},
  year   = {2026},
  url    = {https://huggingface.co/build-small-hackathon/minicpm5-1b-indian-fmcg-normalizer}
}