txn-parser

QLoRA fine-tunes of three small base models for extracting structured transaction data (amount, currency, item, category, type) from free-form Indian-English / code-switched speech and text. Trained on the same 93k-row teacher-labeled dataset, validator-gated, and quantized to 5 GGUF tiers each.

This repo holds all three published students in side-by-side subfolders so a downstream app can pick its size/quality tradeoff:

Base model	Subfolder	Params	Best on-device pick
unsloth/gemma-3-270m-it	gemma-3-270m/	270M	Q5_K_M (260 MB, 99.7% schema)
HuggingFaceTB/SmolLM2-360M-Instruct	smollm2-360m/	360M	Q4_K_M (271 MB, 100% schema)
Qwen/Qwen3-0.6B	qwen3-0.6b/	600M	Q4_K_M (397 MB, 100% schema, best exact match)

Each subfolder contains:

<short>/
├── adapters/                          # PEFT LoRA adapter (load on top of base)
│   ├── adapter_config.json
│   └── adapter_model.safetensors
├── gguf/                              # 5 merged quantizations, ship-ready
│   ├── txn-parser-<short>-F16.gguf
│   ├── txn-parser-<short>-Q8_0.gguf
│   ├── txn-parser-<short>-Q6_K.gguf
│   ├── txn-parser-<short>-Q5_K_M.gguf
│   └── txn-parser-<short>-Q4_K_M.gguf
└── README.md                          # model card with the exact SYSTEM_PROMPT

Eval (300-example held-out set, grammar-constrained decoding)

Model	Quant	Size	JSON valid	Schema valid	Exact match	Amount exact	Mean ms
gemma-3-270m	Q5_K_M	260 MB	99.7%	99.7%	51.0%	84.7%	1788
gemma-3-270m	Q4_K_M	253 MB	93.3%	93.3%	48.3%	80.7%	2660
smollm2-360m	Q5_K_M	290 MB	100.0%	100.0%	52.7%	89.0%	996
smollm2-360m	Q4_K_M	271 MB	100.0%	100.0%	53.3%	87.3%	978
qwen3-0.6b	Q5_K_M	444 MB	100.0%	100.0%	60.0%	91.3%	857
qwen3-0.6b	Q4_K_M	397 MB	100.0%	100.0%	60.0%	90.7%	885

(Showing Q5_K_M + Q4_K_M only — full 15-row table for all 5 quants is in the source repo's eval_results/REPORT.md.)

Recommendations:

• Best accuracy on-device: qwen3-0.6b-Q4_K_M (397 MB, 60% exact, ~885 ms)
• Smallest ship size: smollm2-360m-Q4_K_M (271 MB, 53% exact, ~978 ms)
• Lowest latency: qwen3-0.6b-Q8_0 (639 MB, 59% exact, ~851 ms)
• Avoid: gemma-3-270m-Q4_K_M — quality cliff vs Q5_K_M (93% → 99.7% schema)

Quick download

# Just one quant of one model (small)
huggingface-cli download kartikey31/txn-parser \
    smollm2-360m/gguf/txn-parser-smollm2-360m-Q4_K_M.gguf --local-dir .

# Everything (3 models × 5 quants, ~5 GB)
huggingface-cli download kartikey31/txn-parser --local-dir ./txn-parser

Or from Python:

from huggingface_hub import hf_hub_download

gguf = hf_hub_download(
    "kartikey31/txn-parser",
    "qwen3-0.6b/gguf/txn-parser-qwen3-0.6b-Q4_K_M.gguf",
)

Inference (llama-cpp-python)

from llama_cpp import Llama

# Same SYSTEM_PROMPT for ALL three models — they were trained with it.
SYSTEM_PROMPT = """You convert voice-transcribed transaction descriptions into structured JSON.

Output ONLY a JSON object with this schema, no other text:
{"transactions":[{"amount":<number>,"currency":"INR"|"USD","item":"<lowercase singular noun phrase>","category":"<enum>","type":"expense"|"income"}]}

Categories: Food, Drinks, Groceries, Transport, Shopping, Entertainment, Bills, Health, Education, Personal, Gifts, Income, Other.

Rules:
- Currency defaults to INR. Use USD only when the input explicitly says "dollars" or contains "$".
- Amounts: "k" = ×1000, "hazaar" = ×1000, "sau" = ×100, "lakh" = ×100000. Convert number-words ("five hundred") to digits.
- type is "expense" by default; "income" only for explicit salary, cashback, refund, gift received, payment received.
- For disfluencies and corrections ("500 wait no 600"), output the CORRECTED amount only.
- For ambiguous items ("that thing", "stuff"), use item "unspecified" and category "Other".
- Item field: lowercase singular noun phrase ("uber ride", "beer", "chai" — not "Beers" or "Uber").
- Multi-transaction inputs become multiple array entries in spoken order.
- Category heuristics: uber/ola/auto/petrol/bus/metro → Transport; beer/wine/chai/coffee/juice → Drinks; rent/electricity/wifi/recharge/gas → Bills; movie/netflix/concert → Entertainment; doctor/medicine/hospital → Health."""

llm = Llama(
    model_path="txn-parser-qwen3-0.6b-Q4_K_M.gguf",
    n_gpu_layers=-1, n_ctx=1024,
)
out = llm.create_chat_completion(messages=[
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user",   "content": "200 ka samosa and 50 chai"},
], temperature=0.0)
print(out["choices"][0]["message"]["content"])
# {"transactions":[{"amount":200,"currency":"INR","item":"samosa",...}, ...]}

For Android / on-device deployment guidance (recommended llama.cpp params, battery checklist, Kotlin POC), see the training pipeline README.

Training pipeline

Full reproduction pipeline (data generation → teacher training → distillation → multi-model student training → multi-quant export → eval report) lives at github.com/kartikeychoudhary/txn-parser.

A single command reproduces all three models from a fresh checkout:

git clone https://github.com/kartikeychoudhary/txn-parser
cd txn-parser
bash setup.sh
python scripts/05_generate_distillation_data.py --phase eval --force-eval-copy
python scripts/train_and_publish.py    # trains gemma, smollm, qwen; publishes here
python scripts/eval_all_quants.py      # regenerates the eval table

License

Apache 2.0 (matches all three base model licenses).