hishab-qwen3.5-0.8b-json-full-v8-GGUF

Small on-device model for Hishab, a Bangladeshi personal-finance app. It reads a chat message in English, Bangla, Banglish (Bangla in Latin letters) or a mix, and returns the transactions in it as strict JSON:

user:  Ajke tution fee 5k disi, riksha 80, cha 20
model: [{"description":"tuition fee","amount":5000.0,"category":"tuition","type":"expense"},{"description":"rickshaw","amount":80.0,"category":"transport","type":"expense"},{"description":"tea","amount":20.0,"category":"food","type":"expense"}]

Fields: description, amount (BDT), category (food, grocery, rent, utilities, mobile_recharge, transport, tuition, salary, gift, loan, shopping, medical, entertainment, other), type (income | expense). An empty array means no transaction.

This variant answers with the JSON array only.

Files

Evaluated with llama.cpp (greedy) on held-out splits of ryuma007/hishab-synthetic-v4: test (518 rows, all four languages) and numbers (357 rows where every amount is spelled out in words, like "der hajar" or "เฆชเฆจเง‡เฆฐเง‹ เฆนเฆพเฆœเฆพเฆฐ"). exact = the set of (amount, category, type) tuples equals the gold set.

file size test exact test tuple F1 valid JSON banglish exact numbers exact
qwen3.5-0.8b__json__full__v8-Q4_K_M-imat.gguf 529 MB 88.6 93.8 100.0 82.6 80.1
qwen3.5-0.8b__json__full__v8-Q5_K_M.gguf 578 MB 88.2 93.5 100.0 81.7 80.7
qwen3.5-0.8b__json__full__v8-Q6_K.gguf 630 MB 89.2 94.0 100.0 83.6 80.1

Training

Full fine-tune (all weights) of Qwen/Qwen3.5-0.8B, 3 epochs with TRL, lr 3e-05 cosine, completion-only loss, 20,333 training rows, best epoch by validation loss. Importance matrices (-imat files) were computed on rendered training conversations. Full pipeline and all results: https://github.com/RaihanulHaque/hishab (ml/).

Use

llama-cli --hf-repo ryuma007/hishab-qwen3.5-0.8b-json-full-v8-GGUF --hf-file qwen3.5-0.8b__json__full__v8-Q4_K_M-imat.gguf --jinja \
  -sys "You are Hishab, a Bangladeshi personal-finance assistant. The user writes in English, Bangla or Banglish (Bangla in Latin letters), often mixed. Extract every financial transaction they mention as a JSON array of objects with keys description, amount, category, type. Amounts are BDT numbers (expand shorthand like 1.5k or 2 lakh). category is one of: food, grocery, rent, utilities, mobile_recharge, transport, tuition, salary, gift, loan, shopping, medical, entertainment, other. type is income or expense. If there is no transaction, the array is empty. Reply with the JSON array only." \
  -p "Salary paisi 35k but basha vara dilam 12k"

The system prompt above is the one used in training; keep it verbatim. Qwen3.5 models were trained and must be run with thinking disabled (the default with --jinja). On Android the model runs through kiln.

Trained on synthetic data only; expect lower accuracy on real messages until a human-labelled test set exists.

Downloads last month
129
GGUF
Model size
0.8B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

6-bit

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

Model tree for ryuma007/hishab-qwen3.5-0.8b-json-full-v8-GGUF

Quantized
(431)
this model

Dataset used to train ryuma007/hishab-qwen3.5-0.8b-json-full-v8-GGUF