kartikey31/txn-parser

Transaction Parser — Voice → JSON (on-device)

Distilled student model that turns voice-transcribed transaction strings into structured JSON for an Android expense-tracking app. Examples:

Input	Output
"500 rs on beer 50 rs on candy"	[{amount: 500, item: "beer", category: "Drinks", ...}, {amount: 50, item: "candy", ...}]
"do sau rupay ka chai"	[{amount: 200, currency: "INR", item: "chai", category: "Drinks", ...}]
"1.5k for shoes from myntra"	[{amount: 1500, item: "shoes", category: "Shopping", ...}]
"got my salary 50000"	[{amount: 50000, type: "income", category: "Income", ...}]

What's in this repo

Path	Description
student/gguf/gemma3_text-fixed.BF16.gguf	Lossless ref (543 MB)
student/gguf/gemma3_text-fixed.Q8_0.gguf	High quality (~290 MB)
student/gguf/gemma3_text-fixed.Q5_K_M.gguf	Default for ship (260 MB)
student/gguf/gemma3_text-fixed.Q4_K_M.gguf	Smallest but lossy on 270M (253 MB)
student/adapters/	Trained LoRA adapter (r=32, α=64) for further finetuning
teacher/gguf/gemma-4-e2b-it.Q3_K_M.gguf	Teacher (Gemma 4 E2B) used for distillation labeling
teacher/adapters/	Teacher LoRA adapter (r=16, α=32)

Recommended file

student/gguf/gemma3_text-fixed.Q5_K_M.gguf — 260 MB, 94% JSON valid, runs on-device on Android via llama.cpp at ~150 ms per request on a modern mid-range device.

Evaluation (50-example smoke test)

Build	Size	JSON valid	Schema valid	Exact match (numeric-aware)	Mean latency (A100)
fp16 adapter (ceiling)	n/a	98%	94%	~48%	1219 ms
BF16 GGUF (fixed)	543 MB	98%	74%	48%	108 ms
Q8_0 GGUF (fixed)	~290 MB	~98%	~74%	~46%	~120 ms
Q5_K_M GGUF (fixed)	260 MB	94%	72%	20%	210 ms
Q4_K_M GGUF (fixed)	253 MB	68%	56%	18%	177 ms

The "exact-match" column uses numeric-aware comparison (100 == 100.0). Most "schema invalid" failures are missing-field or enum-value drift; the category prediction is mostly diagonal in the confusion matrix.

Tip for Android: always run a JSON.parse → schema validate → fallback UI pipeline. ~6% of inputs at Q5_K_M will fail to parse — handle that as "couldn't understand, please try again" rather than crashing.

Usage

llama.cpp / llama-cpp-python (Python)

from llama_cpp import Llama

llm = Llama(
    model_path="gemma3_text-fixed.Q5_K_M.gguf",
    n_gpu_layers=-1,
    n_ctx=2048,
    verbose=False,
)

SYSTEM_PROMPT = (
    "You convert short, possibly code-mixed (English/Hindi/Hinglish) "
    "transcribed transaction strings into a JSON object with a single "
    '"transactions" array. Each transaction has: amount (number), '
    "currency (string, default 'INR'), item (string), category (one of "
    "Food, Drinks, Groceries, Transport, Shopping, Entertainment, Bills, "
    "Health, Education, Personal, Gifts, Income, Other), type "
    "('expense' or 'income'). Output ONLY the JSON object — no prose."
)

resp = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": "500 rs on beer 50 rs on candy"},
    ],
    temperature=0.0, top_p=1.0, max_tokens=512,
)
print(resp["choices"][0]["message"]["content"])

Android (llama.cpp JNI)

1. Bundle gemma3_text-fixed.Q5_K_M.gguf in your app (or download on first run).
2. Use the llama.cpp Android example or a JNI wrapper.
3. Set the same system prompt above; user message = the voice transcript.
4. Validate output with a JSON-schema library on the parse path.

Keep the llama_context alive across requests — don't reload per call.

Quick test on Linux/macOS

huggingface-cli download kartikey31/txn-parser \
    --repo-type=model --local-dir models

python -c "
from llama_cpp import Llama
llm = Llama(model_path='models/student/gguf/gemma3_text-fixed.Q5_K_M.gguf', n_gpu_layers=-1, verbose=False)
print(llm.create_chat_completion(messages=[
    {'role':'system','content':'Output only JSON with a transactions array...'},
    {'role':'user','content':'500 rs on beer 50 rs on candy'},
], temperature=0)['choices'][0]['message']['content'])
"

Training details

• Base model: unsloth/gemma-3-270m-it
• Method: QLoRA via Unsloth (r=32, α=64, dropout 0.0, all linear targets)
• Train data: 29,890 teacher-labeled examples (data/distill/train.jsonl) generated by a fine-tuned Gemma 4 E2B teacher
• Epochs: 2
• Effective batch: 128 (A100) / 16 (5060 Ti)
• Optimizer: AdamW 8-bit, cosine LR, peak 2e-4, warmup 3%
• Final eval loss: 0.099 (eval set: 300 hand-curated examples)
• GGUF conversion: raw llama.cpp/convert_hf_to_gguf.py (NOT Unsloth's wrapper), preserves BOS token in chat template
• Hardware: A100-SXM4-80GB, ~25 min total training time at batch 128

Code, dataset generation, evaluation, and conversion scripts: https://github.com/kartikeychoudhary/txn-parser

Categories enum

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

License

Apache-2.0 (matches base model). The training data is synthetic and released under the same license.

Citation

@software{txn-parser-2026,
  author  = {Kartikey Choudhary},
  title   = {Transaction Parser: Voice-to-JSON distilled model},
  year    = {2026},
  url     = {https://huggingface.co/kartikey31/txn-parser},
  note    = {Gemma 3 270M, distilled from Gemma 4 E2B teacher},
}