Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -35,6 +35,9 @@ output in …"). Prompts must follow the VoxSumDroid templates the model was tra
|
|
| 35 |
|
| 36 |
## Training
|
| 37 |
|
|
|
|
|
|
|
|
|
|
| 38 |
Distilled from Qwen3.5-9B over 3,891 meetings (QMSum, MeetingBank, DialogSum,
|
| 39 |
VCSum→OpenCC s2twp; 60k filtered prompt/completion pairs), full-parameter SFT,
|
| 40 |
seq 32768 packed, 2 epochs. Details + eval: training repo `eval/REPORT.md`.
|
|
@@ -44,9 +47,10 @@ seq 32768 packed, 2 epochs. Details + eval: training repo `eval/REPORT.md`.
|
|
| 44 |
| metric | base | fine-tune |
|
| 45 |
|---|---|---|
|
| 46 |
| cross-lingual output-language compliance | 0.38 | **0.98** |
|
| 47 |
-
| action items coverage (1-5) | 1.90 | **3.
|
| 48 |
-
| open questions faith / cover | 3.96 / 2.90 | **4.
|
| 49 |
-
|
|
|
|
|
| 50 |
|
| 51 |
\* base scores "faithful" largely by *not translating* (wrong language 62% of the time).
|
| 52 |
|
|
@@ -54,7 +58,8 @@ seq 32768 packed, 2 epochs. Details + eval: training repo `eval/REPORT.md`.
|
|
| 54 |
|
| 55 |
## Files
|
| 56 |
|
| 57 |
-
- `voxsum-qwen3-0.6b_q8_ekv32768.litertlm` — 664MB, kv 32768
|
| 58 |
-
sha256 `
|
|
|
|
| 59 |
|
| 60 |
RAM note: full 32k KV ≈ 3-3.7GB fp16 on-device — cap `nCtx` lower on small devices.
|
|
|
|
| 35 |
|
| 36 |
## Training
|
| 37 |
|
| 38 |
+
v2 (2026-07-30) adds the single-call structured NOTES task (TITLE/SUMMARY/DECISIONS/
|
| 39 |
+
ACTIONS/OPEN/TOPICS — see docs/OUTPUT-FORMAT.md).
|
| 40 |
+
|
| 41 |
Distilled from Qwen3.5-9B over 3,891 meetings (QMSum, MeetingBank, DialogSum,
|
| 42 |
VCSum→OpenCC s2twp; 60k filtered prompt/completion pairs), full-parameter SFT,
|
| 43 |
seq 32768 packed, 2 epochs. Details + eval: training repo `eval/REPORT.md`.
|
|
|
|
| 47 |
| metric | base | fine-tune |
|
| 48 |
|---|---|---|
|
| 49 |
| cross-lingual output-language compliance | 0.38 | **0.98** |
|
| 50 |
+
| action items coverage (1-5) | 1.90 | **3.94** |
|
| 51 |
+
| open questions faith / cover | 3.96 / 2.90 | **4.42 / 4.10** |
|
| 52 |
+
| structured NOTES format compliance (native / cross) | n/a | **1.00 / 0.94** |
|
| 53 |
+
| known limit: en→zh-TW hour-long faithfulness | 3.73* | ~2.5 |
|
| 54 |
|
| 55 |
\* base scores "faithful" largely by *not translating* (wrong language 62% of the time).
|
| 56 |
|
|
|
|
| 58 |
|
| 59 |
## Files
|
| 60 |
|
| 61 |
+
- `voxsum-qwen3-0.6b_q8_ekv32768.litertlm` — 664MB, kv 32768 (v2)
|
| 62 |
+
sha256 `660436ac918f61f7fddffc419abf561b396585d115382efb8247814ac9382b9a`
|
| 63 |
+
- `checkpoint/` — HF-format safetensors of the same fine-tune (for reuse / reconversion)
|
| 64 |
|
| 65 |
RAM note: full 32k KV ≈ 3-3.7GB fp16 on-device — cap `nCtx` lower on small devices.
|