Qwen3-0.6B fine-tuned for single-pass meeting summarization and insights over VoxSumDroid transcripts, packaged as a LiteRT-LM .litertlm bundle (dynamic int8, 32k context, mask-as-input, Qwen3 ChatML template + stop tokens embedded).

Input format (transcript format v1)

One utterance per line, start-timestamp + speaker tag:

[0:00] S1: 我們今天要決定第三季的行銷預算。
[0:05] S2: 我建議把百分之四十放在社群廣告，效果最好。
[1:23:45] Alice: ...

Variants supported: S1…Sn tags, real names, or no speaker field (no diarization).

Tasks

Summaries (bullet / executive / narrative), title (≤8 words), action items + decisions, detailed summary, topics, open questions, risks/disagreements — in zh-TW and en, including cross-lingual output via the VoxSumDroid language clause ("Write the ENTIRE output in …"). Prompts must follow the VoxSumDroid templates the model was trained on (see the training repo).

Training

v2 (2026-07-30) adds the single-call structured NOTES task (TITLE/SUMMARY/DECISIONS/ ACTIONS/OPEN/TOPICS — see docs/OUTPUT-FORMAT.md).

Distilled from Qwen3.5-9B over 3,891 meetings (QMSum, MeetingBank, DialogSum, VCSum→OpenCC s2twp; 60k filtered prompt/completion pairs), full-parameter SFT, seq 32768 packed, 2 epochs. Details + eval: training repo eval/REPORT.md.

Eval vs base (held-out meetings, teacher-as-judge)

metric	base	fine-tune
cross-lingual output-language compliance	0.38	0.98
action items coverage (1-5)	1.90	3.94
open questions faith / cover	3.96 / 2.90	4.42 / 4.10
structured NOTES format compliance (native / cross)	n/a	1.00 / 0.94
known limit: en→zh-TW hour-long faithfulness	3.73*	~2.5

* base scores "faithful" largely by not translating (wrong language 62% of the time).

Recommended sampler: temp 0.7, top_p 0.8, top_k 20, presence_penalty 1.0.

Files

voxsum-qwen3-0.6b_q8_ekv32768.litertlm — 664MB, kv 32768 (v2) sha256 660436ac918f61f7fddffc419abf561b396585d115382efb8247814ac9382b9a
checkpoint/ — HF-format safetensors of the same fine-tune (for reuse / reconversion)

RAM note: full 32k KV ≈ 3-3.7GB fp16 on-device — cap nCtx lower on small devices.

Luigi/voxsum-qwen3-0.6b-litertlm

voxsum-qwen3-0.6b — on-device meeting summarizer (zh-TW / en)

Input format (transcript format v1)

Tasks

Training

Eval vs base (held-out meetings, teacher-as-judge)

Files

Luigi/
voxsum-qwen3-0.6b-litertlm