HisaabSaathi models

On-device models for HisaabSaathi, an offline voice ledger for daily-wage workers in India.

Understanding model (Qwen3-0.6B, fine-tuned, 8-bit)

hisaabsaathi-qwen3-ft-q8.litertlm: 616,873,984 bytes, sha256 917030d2567e5ddee1ba53c20d95a575f9ca825aebf2a87e71363f099e2290b2.

It does two small jobs. The app does all arithmetic itself.

  1. Given a worker's sentence and the numbers in it, it says what each number is: hours, pieces, rate, wage, received, advance, expense or other.
  2. For a sentence the app's rules did not understand, it says whether it is a question about the worker's records and for which period. The app answers from its own totals.

How it was made:

  • Base model: Qwen/Qwen3-0.6B (Apache-2.0).
  • LoRA fine-tune on both jobs in Hindi, Odia, Telugu, Kannada, Tamil and English.
  • Weights merged, then converted to 8-bit LiteRT-LM with litert-torch.
  • The chat template is Qwen3's, with thinking turned off and no system turn.
  • Runtime: Google LiteRT-LM (Android), CPU.

Choosing it

Qwen3-0.6B, Gemma 3 1B and Qwen2.5-0.5B were fine-tuned the same way on identical data and tested on sentences not used in training.

Held-out sentences labelled right Questions caught (of 147, with the app's rules) Time per answer (PC)
Qwen3-0.6B 140 / 140 118 1.0 s
Gemma 3 1B 127 / 140 100 1.3 s
Qwen2.5-0.5B 127 / 140 54 (the rules alone) 0.8 s
  • After 8-bit conversion: 138 of 140 labelled right, about 1.1 s an answer on a PC.
  • Real speech-recogniser output: 100 of 168 spoken entries fully right with the app's rules, against 95 for the rules alone.
  • No model mistook an ordinary sentence for a question.

English speech model

speech/en-parakeet/: NVIDIA Parakeet TDT-CTC 110M (CC-BY-4.0), int8 ONNX for sherpa-onnx.

Licence

The understanding model is Apache-2.0, the same as its base model. The English speech model keeps its original CC-BY-4.0 licence. Built by Acxelin Labs for HisaabSaathi.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AcxelinLabs/hisaabsaathi-model

Finetuned
Qwen/Qwen3-0.6B
Quantized
(576)
this model