zen-0.6b-v1
On-device task reader for the zen planner app. It turns what a user says or pastes (Vietnamese or English, typos, teencode, self-corrections, scanned notices) into tasks as JSON, with exact dates looked up in a calendar given in the prompt, effort, sittings, parts of the day and a soft deadline.
- Base: Qwen/Qwen3-0.6B with its vocabulary trimmed to Vietnamese and English (151,669 → 45,869 tokens: kept tokens seen at least 20 times in Vietnamese and English Wikipedia samples plus the training inputs, their BPE parents, all 256 byte tokens and the special tokens). Real inputs take 0.5% more tokens than with the full vocabulary.
- Fine-tune: LoRA on all layers (rank 8, 5.0M parameters), 300 steps, fused. Data: about 100 labels written by Qwen3.8-27B with the app's own rules prompt (each with a random "today" and user profile), counted twice, plus 600 calendar drills whose dates were computed by code.
- Files:
model.litertlm: int4 (dynamic_wi4b32_afp32), 4,096-token cache, prefill 256/2048. 280,412,576 bytes, SHA-2563e48aea288f2de62767fbd97198963f2be41ec73f9475f5f6ea15a7d4079e421. This is what the app downloads (expo-ai-kit / LiteRT-LM).hf/: the fused bf16 checkpoint (transformers / mlx-lm).adapter/: the LoRA adapter alone, for the trimmed base in../base/qwen3-0.6b-vien20.
Prompt
ChatML. The chat template accepts content as text or as a list of text parts (LiteRT-LM passes the list form) and ends the generation prompt with an empty think block, as in training. The model then writes the JSON straight away.
- system:
Tách lời nói thành việc. Trả JSON.(documents:Lấy việc của người dùng từ văn bản. Trả JSON.) - user:
Người dùng: <profile>\nBây giờ: YYYY-MM-DD HH:mm.\nLịch:\n<28 lines "YYYY-MM-DD T2 thứ hai">\n\nLời nói:\n<text>(built bymessages()in the app'ssrc/lib/nlu/smart.ts) - answer:
{"tasks":[{"title","type","start","day","deadline","repeat","minutes","sessions","tods","leadDays","dueDays","why"}]}
Held-out results (9 teacher-labelled inputs, small)
| bf16 (mlx) | int4 (LiteRT) | |
|---|---|---|
| Valid JSON | 100% | 100% |
| Task count matches | 89% | about 6 of 9 |
| Dates exact | 6 of 8 | 5 of 9 inputs fully right |
Known misses: lunar dates, all-unaccented teencode titles, an occasional day nobody said (int4). About 12 s per line on a Mac CPU through LiteRT; on-phone speed not measured yet. v2 will train on the full label set (about 1,140).