spent-qwen3-0.6b-expense-parser

SPENT expense parser — Qwen3-0.6B fine-tuned for Swedish (GGUF)

Fine-tuned for the Swedish language. A LoRA fine-tune of Qwen/Qwen3-0.6B for one job: turn a Swedish speech transcript of a purchase into a compact structured line —

"kaffe 39 kronor och bussbiljett 32 kr"  →  [39|SEK|kaffe|Fika;32|SEK|bussbiljett|Transport]
"vi är 3 personer"                        →  []
"fick tillbaka 50 kronor"                 →  []

It understands Swedish amount renderings as spoken and transcribed by Whisper (8 500, 18,90, 12 kronor och 50 öre, en femtiolapp, tusen, typ 200), Swedish currency words (kronor / kr / spänn / bagis, euro, dollar, pund, danska/norska kronor), 300+ Swedish items and merchants, the "two items, one shared price → one row" pattern, up to five purchases in one utterance, and — since v3 — numbers that are not prices (times, counts, ages, bus/phone numbers, durations) and money that is not an expense (refunds, income, loans), which it answers with [].

Runs fully on-device via llama.cpp / llama.rn.

Files

Accuracy — v3 (field accuracy, greedy decoding, training prompt, max_tokens ≥ 160)

Full sweep, all quantization variants, recipes and the full failure analysis: RESULTS.md.

Known limitations (v3)

• Amounts above ~15,000 kr sometimes return [] (bil 125 000 kronor). Training expenses cap at 15,000; large numbers appeared only in income negatives.
• Spelled-out numbers are unreliable (femhundra may parse as 50). Whisper normally normalises numbers to digits, so this rarely reaches the model.
• Income/refund sentences: 4 of 20 still parsed as expenses (skatteåterbäring 640 kronor).
• Smaller quantizations of v3 (313–340 MB) lose 0.1–0.3 pt on ordinary input and, at 313 MB, start refusing real expenses. Ship the 378 MB build.

Usage notes

• Disable Qwen3 thinking (enable_thinking: false / no-think).
• Use the short Swedish system prompt the model was trained on (in RESULTS.md, "For the device session"). A long English prompt still works but costs ~2 pt.
• Allow at least 160 output tokens — five-row answers need ~80.
• Works with a GBNF grammar constraining [amount|CUR|item|Category]; such a grammar forces ≥1 row, so gate silence/empty input before calling the model.
• Output [] means "no expense mentioned".

Alternative: encoder tagger, 120 MB, CC0 (tagger-bert-v2/) — candidate, not yet deployed

A different architecture for the same job: KBLab/bert-base-swedish-cased (CC0-1.0) with a span-tagging head (AMOUNT / ITEM / MERCHANT) and a pointer head that picks the category by reading the category descriptions supplied in the input — so user-added categories work without retraining. Amounts are parsed by a deterministic Swedish number parser, so the output can never be malformed. Runs with ONNX Runtime on CPU.

sha256 tagger-w8.onnx: 30247a2fca44938a250842899a3c544225155d57e1ab9e3877d2dd45474cd7a0. Integration contract (inputs, prompt construction, decoding, pairing rule, parser port): tagger-bert-v2/README.md; reference implementations tagger_common.py, sv_amount.py. Not yet tested on a phone. Full comparison in RESULTS.md.

v3 (2026-09-16) — tagger-bert-v3/, recommended over v2

Same contract and files as v2 (labels.json, vocab.txt, tokenizer.json byte-identical); retrained on data fixed after a 1,191-case iPhone stress run (passive phrasings, hotell, lån till …, öl/tuggummi/tv). 28 remaining phone failures: v2 71.6 % → v3 98.2 %. Held-out sets: private 99.3 → 99.9 %, noisy 95.9 → 99.7 %, buried 87.8 → 96.7 %. tagger-w8.onnx 125,935,273 bytes, sha256 5189611f9c323dda8f96b399a92149f1673209996d11ab409d1b5be1d3f488b6.

Licence

Apache-2.0, inherited from Qwen3-0.6B. Training data is synthetic; no user data.