FinnAI SLM v4

A 1.7B on-device language model for Indian bank SMS → JSON, Ask/Learn finance coaching, and Analytics tips.

Built for FinnDot — private, on-device expense tracking. No user SMS leaves the phone.

Discovery / when to use this model

Use FinnAI SLM v4 if you are building:

• Indian bank SMS → structured JSON (UPI, NEFT, IMPS, CC, salary, ATM)
• On-device / mobile finance assistants (LiteRT-LM INT4 path)
• Ledger-grounded Ask, Learn (tutor), or Analytics tips chat
• Continual fine-tunes from a strong SMS+finance base (adapter/ included)

Related: finndot/finnai-slm-v3, finndot/finnai-slm-v2.

What it does

Languages: English, Hindi, Hinglish, Tamil, Telugu, Marathi, Bengali (+ more Indic mix in train data).

Key results (frozen test)

Protocol: docs/llm-eval-protocol.md. Val SMS R-EM during train: 98.24%.

Quick start (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "finndot/finnai-slm-v4"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="bfloat16", device_map="auto"
)

messages = [
    {
        "role": "system",
        "content": (
            "Extract transaction details from this SMS as JSON: "
            "{amount, merchant, type, account, balance, category}. "
            "type is one of: INCOME, EXPENSE, CREDIT, TRANSFER, INVESTMENT. "
            "If this is not a transaction (OTP, promo, KYC), return {}."
        ),
    },
    {
        "role": "user",
        "content": "HDFC Bank: Rs.499.00 debited from A/c XX1234 to SWIGGY via UPI. Avl Bal Rs.15000.50",
    },
]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Load LoRA-only: adapter/ in this repo (PEFT), base Qwen/Qwen3-1.7B or prior FinnAI merged.

Training details

Privacy: fully synthetic — no user SMS, no real accounts.

On-device (LiteRT-LM)

Production file: finndot/finnai-slm-v4-litertlm

• Format: .litertlm, INT4 dynamic_int4_block32, KV cache 1280, nothink
• Size: 973,979,088 bytes
• SHA256: 36e5ec021730f0e060b78c2371d1e0011c34f28c407cdf9f993232b5cd87e7c0

Limitations

• Optimized for Indian banking SMS formats; weaker on US/EU-only banks (some hard cases remain: FAB, Schwab, etc.).
• Merchant field remains hardest; regex parser-core still preferred when a bank parser exists.
• Chat can still paraphrase; groundedness is improved but not perfect.
• 1.7B — task-specialized, not a general assistant.
• On-device latency gate is measured separately on device hardware.

Reproduce

git clone https://github.com/AnomaProAI/Finndot
cd Finndot/ml
# dataset build + train configs: train/train_config_v4.yaml, data/out_v4_diverse/
python train/sft_qlora.py --config train/train_config_v4.yaml
python -m eval.run_eval --backend hf \
  --sms data/out_v4_diverse/test_sms.jsonl \
  --chat data/out_v4_diverse/chat_eval.jsonl \
  --model-id train/output/v4/merged \
  --baseline-id Qwen/Qwen3-1.7B \
  --tag v20261005-finnai-slm-v4

Citation

@misc{finnai-slm-v4-2026,
  title={FinnAI SLM v4: App-grounded on-device Indian bank SMS parsing and finance coaching},
  author={FinnDot Team},
  year={2026},
  url={https://huggingface.co/finndot/finnai-slm-v4},
  note={Apache 2.0. Continual QLoRA on Qwen3-1.7B / FinnAI v3.}
}

License

Apache 2.0 — same as Qwen3-1.7B.