S1-mini-Chinese-1106-INT8

ไธ“้—จไธบ Qwen3-ASR ไธญๆ–‡ๅŠไธญ่‹ฑๆททๅˆ่ฝฌๅ†™ไผ˜ๅŒ–็š„ๆ–‡ๆœฌๅŽๅค„็†ๆจกๅž‹๏ผŒINT8 W8A16 ็‰ˆๆœฌใ€‚ ๆœฌๆจกๅž‹ๅŸบไบŽ S1-mini by Superwhisper๏ผŒไฟ็•™ๅŽŸๅ็งฐใ€็ฝฒๅใ€LICENSE ๅ’Œ NOTICEใ€‚ ๅฎƒๆฅ่‡ชๅœจไธญๆ–‡ๆ•ฐๆฎไธŠๅ…จ้‡็ปง็ปญ่ฎญ็ปƒ็š„ BF16็จณๅฎšๆ€ง็‰ˆๆœฌ๏ผŒไธๆ˜ฏ้Ÿณ้ข‘่ฏ†ๅˆซๅ™จ๏ผŒไนŸไธๆ˜ฏ้€š็”จ้—ฎ็ญ”ๅŠฉๆ‰‹ใ€‚

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 ผๅผไธŽๅ…ผๅฎนๆ€ง

  • ้€่พ“ๅ‡บ้€š้“ INT8 ๆƒ้‡ใ€BF16 ๆฟ€ๆดปใ€‚

  • ้…็ฝฎๅŒๆ—ถๅฃฐๆ˜Ž rope_theta=1000000 ๅ’Œ rope_parameters.rope_theta=1000000๏ผŒๅ…ผๅฎน Transformers4/5 ็š„ไฝ็ฝฎ็ผ–็ ๅŠ ่ฝฝใ€‚

  • KV cacheใ€embeddingใ€lm_head ๅ’Œๆœช้‡ๅŒ–็š„ๅฝ’ไธ€ๅŒ–้ƒจๅˆ†ไฟ็•™ BF16ใ€‚ๆ ผๅผๆ ‡็ญพไธ่กจ็คบๆฏไธ€ไธชๅผ ้‡้ƒฝๆ˜ฏไฝŽๆฏ”็‰นใ€‚

  • ๅŽ‹็ผฉๅธƒๅฑ€๏ผšcompressed-tensorsใ€‚ๆ–นๆณ•่ฏฆๆƒ…่ง quantization-provenance.jsonใ€‚

  • ๆƒ้‡ๆ–‡ไปถๅˆ่ฎก็บฆ 1.064 GB๏ผ›็”ฑไบŽไฟ็•™ BF16 ๅคง่ฏ่กจๅฑ‚๏ผŒไธ่ƒฝๆŒ‰ๆ ‡็งฐไฝๅฎฝๆŽจๆ–ญๆ•ดไธชๆจกๅž‹ๆœ‰4ๅ€ๅŽ‹็ผฉใ€‚

  • ๆœฌๆฌกๅฎž้™…ๅŠ ่ฝฝๆต‹่ฏ•็กฌไปถไธบ RTX PRO 6000 Blackwell SM120๏ผŒๅŽ็ซฏไธบ vLLM0.17๏ผ›ไธไปฃ่กจๆ‰€ๆœ‰GPUๆˆ–Orin้ƒฝๅฎŒๆˆ้ชŒ่ฏใ€‚NVFP4 ็š„ๅŽŸ็”ŸไฝŽๆฏ”็‰น่ฎก็ฎ—้œ€่ฆ็›ธๅบ”็กฌไปถๆ”ฏๆŒใ€‚

  • ๅชๅ‘ๅธƒๆจกๅž‹ๆ–‡ไปถ๏ผ›ๆœฌๆฌกๅ‘ๅธƒไธๆ›ฟๆข็”จๆˆท็š„็Žฐ็”จ่ฏญ้ŸณๆœๅŠกใ€‚

ๅฎž้™…ๅŽ‹็ผฉๆƒ้‡ๅ›žๅฝ’

ๅŒไธ€ vLLM ้…็ฝฎไธ‹ๆฏ”่พƒ๏ผŒๆฏๆ‰น64ๆกๆไบคใ€ๆœ€ๅคงๅนถๅ‘8ใ€ๆœ€ๅคง็”Ÿๆˆ256 tokensใ€‚่กจไธญBF16ๆˆ็ปฉๆฅ่‡ชไฟฎๆญฃไฝ็ฝฎ็ผ–็ ๅŽ็š„ๅŒ้…็ฝฎ้‡ๆต‹๏ผ›ๆ—ฉๆœŸๅ‘ๅธƒ่ฎฐๅฝ•ไธญ็š„้ป˜่ฎคไฝ็ฝฎ็ผ–็ ๆˆ็ปฉไธไฝœไธบๆœฌๆฌกๅŸบๅ‡†ใ€‚่ฏๆณ•ๅŒน้…ๅŽปๆމ้žๅญ—ๆฏๆ•ฐๅญ—ๅญ—็ฌฆ๏ผŒไฝ†ไฟ็•™ๅคงๅฐๅ†™ไธŽๆ•ฐๅญ—ใ€‚

ๆต‹่ฏ• BF16ๅŸบ็ก€็‰ˆๆœฌ ๆœฌๅŽ‹็ผฉ็‰ˆๆœฌ
ๆ ธๅฟƒๆ‹ผ่ฏป 8/8 8/8
ๅŸบ็ก€่ƒฝๅŠ› 46/48 46/48
ๆ ผๅผไธŽๆ•ฐๅญ— 20/20 20/20
ไธญๆ–‡ไบบๅ 60/60 60/60
ๆ‹ผ่ฏป่ฏญๅขƒ 318/326 318/326
็”Ÿ้€ ่ฏๆ‹ผ่ฏป 196/200 197/200
ๅŸบ็ก€ๆ•ฐๅญฆ 235/236 235/236
ๆ•ฐๅญ—่ฏญๅขƒA 414/432 414/432
ๆ•ฐๅญ—่ฏญๅขƒB 430/432 431/432
ๅ…ฌๅผ€ๆฅๆบๆ–‡ๆœฌ 234/400 237/400

ๅ…ฌๅผ€ๆ–‡ๆœฌ่ฏๆณ• CER๏ผšBF16 9.2367%๏ผŒๆœฌ็‰ˆๆœฌ 9.2257%ใ€‚ ๆœฌๆฌก้ขๅค–ๆ ธ้ชŒๆ‰€้œ€ๆข่กŒใ€ๅๅญ—ๅ†…ๅฎนใ€็ฎ—ๅผ่ฟ็ฎ—็ฌฆๅŠๆ–ฐๅ‡บ็Žฐ็š„ๆ•ฐๅ€ผ้”™่ฏฏ๏ผ›ๅ•่ฏทๆฑ‚ๅ…ณ้”ฎๆฃ€ๆŸฅไธบ 5/5ใ€‚ ่ฟ™ไบ›ๆ˜ฏ้’ˆๅฏนๆ€งๅผ€ๅ‘/ๅ›žๅฝ’ๆฃ€ๆŸฅ๏ผŒไธๆ˜ฏ็‹ฌ็ซ‹็›ฒๆต‹ใ€ไบบๅทฅ่ฏญไน‰ๆญฃ็กฎ็އใ€่‡ช็„ถ่ฏญ้Ÿณ้ชŒๆ”ถๆˆ–ไธŠๆธธ94.8% token accuracy็š„ๅค็Žฐใ€‚ๅฎŒๆ•ด่Œƒๅ›ด่ง evaluation-summary.jsonใ€‚

ไฝฟ็”จ

ๅปบ่ฎฎไฝฟ็”จๆ”ฏๆŒไธŠ่ฟฐๅŽ‹็ผฉๆ ผๅผ็š„ vLLM๏ผŒ้€š่ฟ‡ๆจกๅž‹้…็ฝฎ่‡ชๅŠจ่ฏ†ๅˆซ้‡ๅŒ–ๆ–นๅผใ€‚ไธ่ฆๆŠŠๅŽ‹็ผฉๆƒ้‡ๅฝ“ไฝœๆ™ฎ้€š BF16 safetensors ็›ดๆŽฅๅŠ ่ฝฝใ€‚

vllm serve WIlfLin/S1-mini-Chinese-1106-INT8 --dtype bfloat16 --max-model-len 2048 --max-num-seqs 1

่ฏทๆฑ‚ๅฟ…้กปไฝฟ็”จ BF16ๆจกๅž‹ๅก ไธญ็›ธๅŒ็š„ system prompt ไธŽๆŽงๅˆถ่กŒ๏ผŒๅนถๅ…ณ้—ญ thinkingใ€‚ๆ™ฎ้€šๆ–‡ๅญ—้€‰ๆ‹ฉ Structure: prose๏ผŒๆ˜Ž็กฎๅˆ—่กจ้€‰ๆ‹ฉ Structure: listsใ€‚ๆจกๅž‹่‡ช่กŒ็”Ÿๆˆๆ ‡็‚นใ€็ผ–ๅทไธŽๆข่กŒ๏ผ›ไธ้™„ๅธฆๆ‹ผ่ฏปๅ‰็ฝฎ่ง„ๅˆ™ๆˆ–ๅ†…ๅฎนไฟๆŠค่ฟ‡ๆปคๅ™จใ€‚ๅบ”็”จ่ดŸ่ดฃ่ฏทๆฑ‚ๅผ‚ๅธธใ€ๆˆชๆ–ญๆฃ€ๆต‹๏ผŒไปฅๅŠๅฎ‰ๅ…จๆไบคๆข่กŒใ€‚

่ฎญ็ปƒไธŽ้™ๅˆถ

ๅŸบ็ก€็‰ˆๆœฌ็ป่ฟ‡ๅ…จๅ‚ๆ•ฐไธญๆ–‡่ฎญ็ปƒใ€ๆ•ฐๅ€ผ็ป„ๅˆๅ›žๆ”พใ€ๅฏไฟกFP32ๆ•™ๅธˆ็š„ไธ€่‡ดๆ€ง็บฆๆŸๅŠๅ…ณ้”ฎๅญ—็ฌฆๅˆ†ๆ•ฐ้—ด้š”่ฎญ็ปƒใ€‚่ฏฆๆƒ…่ง training-provenance.jsonใ€data-provenance.json ๅ’Œ BF16 ๆจกๅž‹ๅกใ€‚ๆ ‡ๅฎšๅชไฝฟ็”จ่ฎญ็ปƒๆ ทๆœฌ๏ผ›ๆœฌไป“ๅบ“ไธๅ‘ๅธƒๆ•ฐๆฎๆญฃๆ–‡ใ€็งๆœ‰ๅฝ•้Ÿณๆˆ–็”จๆˆท่พ“ๅ…ฅๅކๅฒใ€‚

้‡ๅŒ–ไธไผšๆถˆ้™คๅŸบ็ก€ๆจกๅž‹็š„้”™่ฏฏใ€‚ๆ•ฐๅญ—ๅฏ่ƒฝๅคš่กฅ้›ถ๏ผŒไธ“ๆœ‰่ฏใ€ๆ‹ผ่ฏปๅคงๅฐๅ†™ใ€ๅฆๅฎšใ€ๆ”นๅฃไธŽ้•ฟๆฎต่ฝไปๅฏ่ƒฝๅ‡บ้”™ใ€‚ไธๅŒๅŽ็ซฏใ€ๆ‰นๆฌกๅ’Œ็กฌไปถๅฏ่ƒฝๆ”นๅ˜็ป“ๆžœ๏ผ›ไธŠ่ฟฐๆฃ€ๆŸฅไธไฟ่ฏไปปๆ„่พ“ๅ…ฅ้ƒฝๆญฃ็กฎใ€‚ๅ„็ฒพๅบฆ็‰ˆๆœฌๅฟ…้กปๆ นๆฎ่‡ชๅทฑ็š„ๅ›žๅฝ’ๆˆ็ปฉ้€‰ๆ‹ฉ๏ผŒไธ่ƒฝ็›ดๆŽฅๆฒฟ็”จ BF16 ๆˆ็ปฉใ€‚

่ฎธๅฏ่ฏไธŽ็ฝฒๅ

ๆœฌๆจกๅž‹ๆ˜ฏ S1-mini by Superwhisper ็š„ไฟฎๆ”น็‰ˆๆœฌ๏ผŒๅบ•ๅฑ‚ๆžถๆž„ไธบ Qwen3-0.6Bใ€‚ไฟ็•™ๅŽŸๅง‹ LICENSE ไธŽ NOTICE๏ผŒๅŒ…ๆ‹ฌ็ปง็ปญไฟ็•™ๅŽŸๅ็งฐๅ’Œไฝœ่€…็ฝฒๅ็š„้™„ๅŠ ๆกๆฌพใ€‚CS2W ๅฃฐๆ˜Ž CC-BY-SA-4.0๏ผŒDisfl-QA ๅฃฐๆ˜Ž CC-BY-4.0๏ผ›ไธๅฐ†ๆ•ฐๆฎๆกๆฌพๆ”น็งฐไธบ็บฏ Apache-2.0ใ€‚

Downloads last month
30
Safetensors
Model size
0.8B params
Tensor type
I32
ยท
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for WIlfLin/S1-mini-Chinese-1106-INT8

Finetuned
Qwen/Qwen3-0.6B
Quantized
(2)
this model