S1-mini-Chinese-1106-INT8
ไธ้จไธบ Qwen3-ASR ไธญๆๅไธญ่ฑๆททๅ่ฝฌๅไผๅ็ๆๆฌๅๅค็ๆจกๅ๏ผINT8 W8A16 ็ๆฌใ ๆฌๆจกๅๅบไบ S1-mini by Superwhisper๏ผไฟ็ๅๅ็งฐใ็ฝฒๅใLICENSE ๅ NOTICEใ ๅฎๆฅ่ชๅจไธญๆๆฐๆฎไธๅ จ้็ปง็ปญ่ฎญ็ป็ BF16็จณๅฎๆง็ๆฌ๏ผไธๆฏ้ณ้ข่ฏๅซๅจ๏ผไนไธๆฏ้็จ้ฎ็ญๅฉๆใ
Chinese-data continued fine-tuning of S1-mini by Superwhisper, optimized for Qwen3-ASR Chinese and mixed Chinese/English transcript postprocessing. This repository contains actual INT8 compressed weights, not renamed BF16 files.
ๆ ผๅผไธๅ ผๅฎนๆง
้่พๅบ้้ INT8 ๆ้ใBF16 ๆฟๆดปใ
้ ็ฝฎๅๆถๅฃฐๆ
rope_theta=1000000ๅrope_parameters.rope_theta=1000000๏ผๅ ผๅฎน Transformers4/5 ็ไฝ็ฝฎ็ผ็ ๅ ่ฝฝใKV cacheใembeddingใlm_head ๅๆช้ๅ็ๅฝไธๅ้จๅไฟ็ BF16ใๆ ผๅผๆ ็ญพไธ่กจ็คบๆฏไธไธชๅผ ้้ฝๆฏไฝๆฏ็นใ
ๅ็ผฉๅธๅฑ๏ผ
compressed-tensorsใๆนๆณ่ฏฆๆ ่ง quantization-provenance.jsonใๆ้ๆไปถๅ่ฎก็บฆ 1.064 GB๏ผ็ฑไบไฟ็ BF16 ๅคง่ฏ่กจๅฑ๏ผไธ่ฝๆๆ ็งฐไฝๅฎฝๆจๆญๆดไธชๆจกๅๆ4ๅๅ็ผฉใ
ๆฌๆฌกๅฎ้ ๅ ่ฝฝๆต่ฏ็กฌไปถไธบ RTX PRO 6000 Blackwell SM120๏ผๅ็ซฏไธบ vLLM0.17๏ผไธไปฃ่กจๆๆGPUๆOrin้ฝๅฎๆ้ช่ฏใNVFP4 ็ๅ็ไฝๆฏ็น่ฎก็ฎ้่ฆ็ธๅบ็กฌไปถๆฏๆใ
ๅชๅๅธๆจกๅๆไปถ๏ผๆฌๆฌกๅๅธไธๆฟๆข็จๆท็็ฐ็จ่ฏญ้ณๆๅกใ
ๅฎ้ ๅ็ผฉๆ้ๅๅฝ
ๅไธ vLLM ้ ็ฝฎไธๆฏ่พ๏ผๆฏๆน64ๆกๆไบคใๆๅคงๅนถๅ8ใๆๅคง็ๆ256 tokensใ่กจไธญBF16ๆ็ปฉๆฅ่ชไฟฎๆญฃไฝ็ฝฎ็ผ็ ๅ็ๅ้ ็ฝฎ้ๆต๏ผๆฉๆๅๅธ่ฎฐๅฝไธญ็้ป่ฎคไฝ็ฝฎ็ผ็ ๆ็ปฉไธไฝไธบๆฌๆฌกๅบๅใ่ฏๆณๅน้ ๅปๆ้ๅญๆฏๆฐๅญๅญ็ฌฆ๏ผไฝไฟ็ๅคงๅฐๅไธๆฐๅญใ
| ๆต่ฏ | BF16ๅบ็ก็ๆฌ | ๆฌๅ็ผฉ็ๆฌ |
|---|---|---|
| ๆ ธๅฟๆผ่ฏป | 8/8 | 8/8 |
| ๅบ็ก่ฝๅ | 46/48 | 46/48 |
| ๆ ผๅผไธๆฐๅญ | 20/20 | 20/20 |
| ไธญๆไบบๅ | 60/60 | 60/60 |
| ๆผ่ฏป่ฏญๅข | 318/326 | 318/326 |
| ็้ ่ฏๆผ่ฏป | 196/200 | 197/200 |
| ๅบ็กๆฐๅญฆ | 235/236 | 235/236 |
| ๆฐๅญ่ฏญๅขA | 414/432 | 414/432 |
| ๆฐๅญ่ฏญๅขB | 430/432 | 431/432 |
| ๅ ฌๅผๆฅๆบๆๆฌ | 234/400 | 237/400 |
ๅ ฌๅผๆๆฌ่ฏๆณ CER๏ผBF16 9.2367%๏ผๆฌ็ๆฌ 9.2257%ใ ๆฌๆฌก้ขๅคๆ ธ้ชๆ้ๆข่กใๅๅญๅ ๅฎนใ็ฎๅผ่ฟ็ฎ็ฌฆๅๆฐๅบ็ฐ็ๆฐๅผ้่ฏฏ๏ผๅ่ฏทๆฑๅ ณ้ฎๆฃๆฅไธบ 5/5ใ ่ฟไบๆฏ้ๅฏนๆงๅผๅ/ๅๅฝๆฃๆฅ๏ผไธๆฏ็ฌ็ซ็ฒๆตใไบบๅทฅ่ฏญไนๆญฃ็กฎ็ใ่ช็ถ่ฏญ้ณ้ชๆถๆไธๆธธ94.8% token accuracy็ๅค็ฐใๅฎๆด่ๅด่ง evaluation-summary.jsonใ
ไฝฟ็จ
ๅปบ่ฎฎไฝฟ็จๆฏๆไธ่ฟฐๅ็ผฉๆ ผๅผ็ vLLM๏ผ้่ฟๆจกๅ้ ็ฝฎ่ชๅจ่ฏๅซ้ๅๆนๅผใไธ่ฆๆๅ็ผฉๆ้ๅฝไฝๆฎ้ BF16 safetensors ็ดๆฅๅ ่ฝฝใ
vllm serve WIlfLin/S1-mini-Chinese-1106-INT8 --dtype bfloat16 --max-model-len 2048 --max-num-seqs 1
่ฏทๆฑๅฟ
้กปไฝฟ็จ BF16ๆจกๅๅก ไธญ็ธๅ็ system prompt ไธๆงๅถ่ก๏ผๅนถๅ
ณ้ญ thinkingใๆฎ้ๆๅญ้ๆฉ Structure: prose๏ผๆ็กฎๅ่กจ้ๆฉ Structure: listsใๆจกๅ่ช่ก็ๆๆ ็นใ็ผๅทไธๆข่ก๏ผไธ้ๅธฆๆผ่ฏปๅ็ฝฎ่งๅๆๅ
ๅฎนไฟๆค่ฟๆปคๅจใๅบ็จ่ด่ดฃ่ฏทๆฑๅผๅธธใๆชๆญๆฃๆต๏ผไปฅๅๅฎๅ
จๆไบคๆข่กใ
่ฎญ็ปไธ้ๅถ
ๅบ็ก็ๆฌ็ป่ฟๅ จๅๆฐไธญๆ่ฎญ็ปใๆฐๅผ็ปๅๅๆพใๅฏไฟกFP32ๆๅธ็ไธ่ดๆง็บฆๆๅๅ ณ้ฎๅญ็ฌฆๅๆฐ้ด้่ฎญ็ปใ่ฏฆๆ ่ง training-provenance.jsonใdata-provenance.json ๅ BF16 ๆจกๅๅกใๆ ๅฎๅชไฝฟ็จ่ฎญ็ปๆ ทๆฌ๏ผๆฌไปๅบไธๅๅธๆฐๆฎๆญฃๆใ็งๆๅฝ้ณๆ็จๆท่พๅ ฅๅๅฒใ
้ๅไธไผๆถ้คๅบ็กๆจกๅ็้่ฏฏใๆฐๅญๅฏ่ฝๅค่กฅ้ถ๏ผไธๆ่ฏใๆผ่ฏปๅคงๅฐๅใๅฆๅฎใๆนๅฃไธ้ฟๆฎต่ฝไปๅฏ่ฝๅบ้ใไธๅๅ็ซฏใๆนๆฌกๅ็กฌไปถๅฏ่ฝๆนๅ็ปๆ๏ผไธ่ฟฐๆฃๆฅไธไฟ่ฏไปปๆ่พๅ ฅ้ฝๆญฃ็กฎใๅ็ฒพๅบฆ็ๆฌๅฟ ้กปๆ นๆฎ่ชๅทฑ็ๅๅฝๆ็ปฉ้ๆฉ๏ผไธ่ฝ็ดๆฅๆฒฟ็จ BF16 ๆ็ปฉใ
่ฎธๅฏ่ฏไธ็ฝฒๅ
ๆฌๆจกๅๆฏ S1-mini by Superwhisper ็ไฟฎๆน็ๆฌ๏ผๅบๅฑๆถๆไธบ Qwen3-0.6Bใไฟ็ๅๅง LICENSE ไธ NOTICE๏ผๅ ๆฌ็ปง็ปญไฟ็ๅๅ็งฐๅไฝ่ ็ฝฒๅ็้ๅ ๆกๆฌพใCS2W ๅฃฐๆ CC-BY-SA-4.0๏ผDisfl-QA ๅฃฐๆ CC-BY-4.0๏ผไธๅฐๆฐๆฎๆกๆฌพๆน็งฐไธบ็บฏ Apache-2.0ใ
- Downloads last month
- 30
Model tree for WIlfLin/S1-mini-Chinese-1106-INT8
Base model
Qwen/Qwen3-0.6B-Base