Qwen3.5-0.8B FT (Q4_K_M) — Workvoice on-device extraction
Fine-tuned small language model used by Workvoice for offline structured
field extraction from Polish HVAC / fire-safety inspection transcripts, running
fully on-device via llama.rn (llama.cpp).
- File:
Qwen3.5-0.8B-FT-Q4_K_M.gguf (~505 MB, Q4_K_M quantization)
- Task: given a field descriptor + a Polish voice transcript, return the field value
(GBNF-constrained decoding, reasoning disabled).
- Training data: synthetic Polish inspection cases only — no customer data.
- On-device bench: ~60% strict exact-match on the 114-case suite; ~475 ms/field
on a mid-range 2023 Android device.
Intended use
Downloaded once on demand by the Workvoice mobile app and cached on device; used only
when the device is offline. When online, the app uses its cloud extraction path.
Notes
- Requires a runtime supporting range requests for resumable download (the HF CDN does).
- Not intended as a general-purpose chat model.
