S1-mini-Chinese-1106
ไธ้จไธบ Qwen3-ASR ไธญๆๅไธญ่ฑๆททๅ่ฝฌๅไผๅ็ๆๆฌๅๅค็ๆจกๅใ
ๆฌๆจกๅๆฏๅจไธญๆๅฃ่ฏญๆด็ๆฐๆฎไธๅๆไฟฎ่ฎขๆฐๆฎไธ็ปง็ปญๅ
จ้ๅพฎ่ฐ็็ๆฌใ
ๅบไบ S1-mini by Superwhisper๏ผไฟ็ๅๅงๅ็งฐใ็ฝฒๅใLICENSE ๅ NOTICEใ
ๆฌๆฌกๅๅธๅฏนๅบ native-v10 ๆ็ป BF16 ๆ้๏ผไธ 2026-11-06 ้จ็ฝฒ็ๆฌไธ่ดใ
A Chinese-data full fine-tune of S1-mini by Superwhisper, optimized for postprocessing Qwen3-ASR Chinese and Chinese/English mixed transcripts. This is a text postprocessor, not an audio recognizer or a general-purpose assistant. Release 1106 corresponds to the native-v10 final checkpoint.
ๆฌๆฌกๆดๆฐ
- ๆจกๅๅญฆไน ่ฑๆๆผ่ฏปๅๅนถไธไนฆ้ขๅคงๅฐๅ๏ผ็ฉบๆ ผๅญๆฏใ่ฟๅญ็ฌฆๅญๆฏใ่ฟ็ปญๅคงๅใ่ฟ้ณ่ฏๅๆผ่ฏป๏ผไปฅๅๆ็กฎๅฑ็คบๆผๅ็่ฏญๅขใ
- ๆ็กฎๅฑ็คบๆผๅๆถไฝฟ็จๅญๆฏ้ด่ฟๅญ็ฌฆ๏ผๆฎ้่ฏใๅ ฌๅธใไบบๅๅ็ผฉๅๆ่ฏญๅขไนฆๅใ
- ๆจกๅ็ๆๆฎต่ฝใๆข่กๅ็ผๅทๅ่กจ๏ผๆ ้ๅค้จๆ ผๅผๆด็่งๅใ
- ๅขๅ ไธญๆๅ่ฑๆๅบ็กๆฐๅญฆ่ฝฌๅ๏ผ
+ใ-ใ*ใ/ใ^ใ=๏ผไธ่พๅบ LaTeXใไธไธปๅจ่ฎก็ฎๆไฟฎๆญฃๅฃ่ฟฐ็ญๆกใ - ๅขๅ ๆฐๅญๅ ้จ้ถใๆถ้ด้ถๅ้ใไธญๆไบบๅไฟ็ใๅฆๅฎใๆนๅฃๅไธๅฎๆด็ๆฎต่ฎญ็ปใ
- ไฟ็ๅปๅฃๆฐด่ฏใๅป้ๅคๅๅๆไฟๆ่ฝๅ็ๅๆพใ
่ฟไบๆฏ่ฎญ็ป็ฎๆ ๏ผไธๆฏๆๆ่พๅ ฅไธ็ไฟ่ฏใ้จๅๅ่กจ่ฏญๅขใไธไธ่ฏใๆฐๅญๅ้ฟ่ท็ฆปๆนๅฃไปไผๅบ้ใ ๆผ่ฏปใๆฐๅญฆๅๆ ผๅผ่ฝๅๆฅ่ชๆจกๅ๏ผๆฌไปๅบไธ้ๅธฆๆง็ๆผ่ฏปๅ็ฝฎ่งๅๆๅ ๅฎนไฟๆค่ฟๆปคๅจใ ่ฐ็จๆนไป้กปๅค็่ถ ๆถใ็ๆๆชๆญๅๆๆฌๆไบคใๆจกๅ็ๆๆข่กไธไผ่ชๅจไฟ่ฏ่พๅ ฅๆณไธ่งฆๅๅ้้ฎใ
่ฎญ็ปๆนๆณไธๆฅๆบ
็บฆ 7.52 ไบฟๅๆฐ๏ผๅ จๅๆฐ็็ฃๅพฎ่ฐ๏ผไธๆฏ LoRA๏ผไนไธๆฏๅช่งฃๅปๆๅๅ ๅฑใ ็ปงๆฟๅๆไธญๆ่ฎญ็ป๏ผ้ๅๅ ๅ ฅๅ่ฝๅๅๆพใๆผ่ฏป่ฏญๅขใๆฐๅญฆๅๅญๆฎตไฟ็่ฏพ็จใ ไธญๆๅบ็กๆฅๆบไธบ CS2W๏ผ่ฑๆๆนๅฃๆฅๆบไธบ Disfl-QA๏ผๅฆๆ่ชๅถๅๆจกๅ่พ ๅฉๅๆๆฐๆฎใ ๅๆๆ ๆณจ็ป่ฟ็จๅบๆ ธ้ช๏ผ่ฟไธ็ญไบ็ฌ็ซไบบๅทฅๆ ๆณจ๏ผไนไธ็ญไบ่ช็ถ่ฏญ้ณๆต่ฏใ
ๆ่ฟ้ถๆฎต๏ผ
| ้ถๆฎต | ๆ ทๆฌๅ็ฐๆฑ | ้ๅ็ปๆ |
|---|---|---|
| v8.2 ็ปผๅ่ฏพ็จ | 21,240 | ไธ่ฝฎ๏ผ1,328 ๆญฅ๏ผๅญฆไน ็ 2e-6 |
| v9 ้่ฏฏไฟฎ่ฎขไธๅๆพ | 26,886 | ้ๅ็ฌฌ 900 ๆญฅ๏ผๅญฆไน ็ 1e-6 |
| v10 ๅญๆฎตไฟ็็ฒพไฟฎ | 3,104 | ไธ่ฝฎ๏ผ194 ๆญฅ๏ผๅญฆไน ็ 5e-7 |
ๆ็ป้ถๆฎตๅ
ๅซ 1,494 ๆกๆฐๅขๅญๆฎต่ฏพ็จๅ 1,610 ๆกๅๆพ๏ผ12 ๆกไธบ็ฉบ่พๅบ็็ฃใ
้ๅค้ๆ ท่กจ็คบๆ้๏ผไธ่กจ็คบๅ็ญๆฐ้็็ฌ็ซๆ ทๆฌใ่ฎญ็ปไป
ๅฏน็ฎๆ ่พๅบไธ EOS ่ฎก็ฎๆๅคฑใ
ๆฐๅขๆผ่ฏป็ฎๆ ้็จๆ ่ฎฐ็ผ็ ๏ผๆญฃๅธธ่ฐ็จไป็ดๆฅไฝฟ็จ้ๅธฆ tokenizer ๅ chat templateใ
่ฎญ็ป้็จ FP32 ไธปๅๆฐ/ไผๅๅจไธ BF16 autocast๏ผๅฏผๅบๆ้ไธบ BF16ใ
ไธๆธธๅบๅฎ็ๆฌ๏ผsuperwhisper/s1-mini@88f6b15896c73bbb13a3b596e0afe8ea0d5150b4ใ
่ง training-provenance.json ๅ data-provenance.jsonใๆฌๅๅธไธๅ ๅซ่ฎญ็ปๆๆฌใ็งๆๅฝ้ณใ็จๆท่พๅ ฅๅๅฒๆ้จ็ฝฒ้ ็ฝฎใ
ๆต่ฏ็ปๆไธ้ๅถ
ไปฅไธไธบ TransformersใBF16ใgreedyใbatch 16ใๆๅคง็ๆ 256 tokens ็ๅๅงๆจกๅ่พๅบ๏ผๆฒกๆๅค้จๅ ๅฎนๆ ก้ชใ
| ๅผๅๅๅฝ้ | ่ฏๆณๆดๅฅๅน้ | ไธฅๆ ผๆดๅฅๅน้ |
|---|---|---|
| ๅบ็ก่ฝๅ | 46/48 | 42/48 |
| ๅ่กจไธๆฐๅญๆ ผๅผ | 20/20 | 19/20 |
| ไธญๆไบบๅไฟ็ | 60/60 | 60/60 |
| ๆผ่ฏป่ฏญๅข | 318/326 | 318/326 |
| ็้ ่ฏๆผ่ฏปไฟ็ | 197/200 | 197/200 |
| ๅบ็กๆฐๅญฆ่ฝฌๅ | 235/236 | 235/236 |
| ไฟฎ่ฎขๅผๅ้ | 81/94 | 80/94 |
| ไฟฎ่ฎข่ฏๆญ้ | 74/97 | 73/97 |
| ๅ ฌๅผๆฅๆบๆๆฌๅๅฝ | 232/400 | 118/400 |
ๅ ฌๅผๆๆฌๅๅฝ็่ฏๆณๅญ็ฌฆ้่ฏฏ็ไธบ **9.3685%**ใ่ฏๆณๅน้ ๅ ้ค้ๅญๆฏๆฐๅญๅญ็ฌฆ๏ผไฟ็ๅญๆฏๅคงๅฐๅไธๆฐๅญ๏ผไธฅๆ ผๅน้ ไฟ็ๆ ็นๅๅ ้จ็ฉบ็ฝใ่ฟไบๆฏๅ่ๆๆฌๅน้ ๆๆ ๏ผไธๆฏไบบๅทฅ่ฏญไนๆญฃ็กฎ็ใ
่ฟไบ้ๅๅจๅผๅไธญๅๅคไฝฟ็จ๏ผไธๆฏๆฐ็็ฌ็ซ็ฒๆตใไบบๅ 60/60 ไธไปฃ่กจไปปๆไบบๅ้ฝๆญฃ็กฎใ็ธๆฏ v9 ็ฌฌ 900 ๆญฅ๏ผๆฌ็ๆนๅไบบๅ๏ผไฝ 97 ๆก่ฏๆญ้็ๆปๅไธ้๏ผไธ่ฝๅฃฐ็งฐๆๆ่ฝๅๅๆ้ซใ
BF16 ๅฏนๅ็ซฏใ็ฒพๅบฆๅๆนๆฌกๅฝข็ถๆๆใ ๅไธๅฅ 501+3=504 ่ฝฌๅๅจไธไธช batch 16 ไธญ้่ฏฏ่พๅบ 5001+3=504๏ผ้ๆก็ๆๅ FP32 ๅฏน็
ง่พๅบๆญฃ็กฎใ97 ๆก่ฏๆญ้้ๆก BF16 ไธ batch 16 FP32 ๅไธบ 75/97 ไธฅๆ ผๅน้
ใFP32 ๅนถ้ๆๆ้่ฏฏ็้็จไฟฎๅคใ
Orin vLLM 0.23.0ใBF16ใๅๅบๅๆๅก้่ฟ 23 ๆก้ๅฏนๆงๆจกๅๆฃๆฅๅไธๆก็ฆ็จๆด็็็ด้ๆฃๆฅใ่ฟไป ่ฏๆๆๅฎๅบๆฏ็็ๅฎๆๅก่กไธบ๏ผไธๆฏๅนฟๆณๆญฃ็กฎ็ๆ็ๅฎๅฝ้ณ้ชๆถใ
ๆฒกๆไธบๅฝๅๆ้้็ฎไธๆธธๆ็งฐ็ 94.8% token accuracy๏ผไธๅพๅฐไธ่ฟฐๆ็ปฉ่งไธบ่ฏฅๅฃๅพๅค็ฐใ ๅฎๆดๆฑๆป่ง evaluation-summary.jsonใ
ๆจ็็คบไพ
ไฝฟ็จ้ๅธฆๆจกๆฟ๏ผๅ
ณ้ญ thinkingใๆฎ้ๆๆฌ้ๆฉ Structure: prose๏ผๆ็กฎๅ่กจ้ๆฉ Structure: listsใ
ไปฅไธ็คบไพ็ดๆฅ่ฟๅๆจกๅ่พๅบ๏ผไธ่ฐ็จๆผ่ฏป่งๅๆๆ ผๅผๅๅค็ใ
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "WIlfLin/S1-mini-Chinese-1106"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id, dtype=torch.bfloat16, device_map="auto"
).eval()
messages = [
{"role": "system", "content": (
"You are a text normalizer for speech-to-text transcripts. The input begins "
"with a control line specifying the styling, structure, and context settings; "
"clean the transcript to match those settings and output only the cleaned text."
)},
{"role": "user", "content": (
"[Styling: semi-formal] [Structure: lists] [Context: general]\n"
"ไปๅคฉๆไธคไปถไบ๏ผ็ฌฌไธๆฃๆฅๅคไปฝ๏ผ็ฌฌไบๆดๆฐๆๆกฃใ"
)},
]
text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.inference_mode():
generated = model.generate(
**inputs, do_sample=False, max_new_tokens=256,
pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(
generated[0, inputs.input_ids.shape[1]:], skip_special_tokens=True
).strip())
ๅชๅซๅฃๆฐด่ฏๆถๅฏไปฅ่พๅบ็ฉบๅญ็ฌฆไธฒใ่พๅ ฅไธญ็ๆ้ฎใๅฝไปคใๅผ็จๅๆฏๅพ ๆด็ๆๆฌ๏ผๆจกๅไธๅบๅ็ญๆๆง่กใๅฎ้ ่พๅบไปๅฏ่ฝๅ็ฆป่ฟไบ็ฎๆ ใ้ฟๆๆฌๅบๅจๅบ็จไธญๆ็กฎๆฃๆตๆชๆญ๏ผไธ่ฝๅฐๆชๆญ็ปๆๅฝไฝๅฎๆดๆด็ใ
่ฎธๅฏ่ฏไธ็ฝฒๅ
ๆฌๆจกๅๆฏ S1-mini by Superwhisper ็ไฟฎๆน็ๆฌ๏ผๅบๅฑๆถๆไธบ Qwen3-0.6Bใ ไฟ็ๅๅง LICENSE ๅ NOTICEใๅ่ฎธๅฏ่ฏๅบไบ Apache-2.0๏ผๅนถๅซ็ปง็ปญไฟ็ๅๅงๅ็งฐไธไฝ่ ็ฝฒๅ็้ๅ ๆกๆฌพ๏ผไธ่ฝ็็ฅ่ฏฅๆกๆฌพใ ๆฐๆฎๆฅๆบๅฃฐๆ๏ผCS2W ไธบ CC-BY-SA-4.0๏ผDisfl-QA ไธบ CC-BY-4.0๏ผๆฐๆฎๆกๆฌพ่งๆฅๆบๆไปถ๏ผไธๆน็งฐไธบ็บฏ Apache-2.0ใ
ๆ้ SHA256๏ผ79d38b1e577677c88d911d70d20ad110a46b54426a7c87f0cb7433c1133575f9ใ
- Downloads last month
- -