S1-mini-Chinese
ๆฌๆจกๅไธ้จ้ขๅ Qwen3-ASR ็ไธญๆๅไธญ่ฑๆททๅ่ฝฌๅๅๅค็ไผๅใ
่ฟๆฏๅจไธญๆๅฃ่ฏญๆด็ๆฐๆฎ้ไธ็ปง็ปญๅพฎ่ฐ็ S1-mini ็ๆฌใ
ๆฌๆจกๅๅบไบ S1-mini by Superwhisper๏ผไฟ็ๅๅงๆจกๅๅ็งฐใไฝ่
็ฝฒๅใLICENSE ๅ NOTICEใ
ๅฝๅ็ๆฌๆฏ format-numbers-v5๏ผ2026-11-05๏ผ๏ผไธ็ฎๅ้จ็ฝฒ็ BF16 ๆ้ไธ่ดใ
Chinese-data continued full fine-tuning of S1-mini by Superwhisper, optimized for postprocessing Qwen3-ASR transcripts in Chinese and mixed Chinese/English. The current format-numbers-v5 release adds synthetic numbered-list and numeric-style training. This is a text postprocessor, not a speech recognizer.
ๅฝๅ็ๆฌ็ๆนๅจ
- ๆ็กฎๅไธพไฝฟ็จ
1.ใ2.ใ3.ๆ็ผๅทๅ่กจ๏ผๆฏ้กนๅ็ฌๆข่กใ - ๅขๅ ้ๅทใๅฅๅทใๅๅทๅ้็ๅไธพ๏ผไปฅๅ่พ้ฟๅ่กจ้กน่ฎญ็ป๏ผๅๅฐๅธฆๆ ็นๆถ็ งๆๅๆ็้ฎ้ขใ
- ๅๅทใ็ๆฌๅทใๅๆฐ้ไผๅ
ไฝฟ็จ้ฟๆไผฏๆฐๅญ๏ผไพๅฆ
iPhone 18ใQwen2.5 7Bใ - ๅคงๆดๆฐไฟ็ไธใไบฟ็ไธญๆๅไฝๅๆณ๏ผๆ่ฏญๅๅบๅฎ่ฏดๆณไฟ็ไธญๆ๏ผ้ช่ฏ็ ่ฎญ็ปๅ ๅซๅๅฏผ้ถใ
- ไฟ็ๅปๅฃๆฐด่ฏใ้ๅคใๆ็กฎ่ชๆ็บ ๆญฃๅไธไฟฎๆนๆ ทๆฌ็ๅๆพใ
่ฟไบๆฏ่ฎญ็ป็ฎๆ ๏ผไธๆฏๆๆ่พๅ ฅไธ็ไฟ่ฏใๅคๆๆฐๅญ่ฟ่ฏปใ่ฏๅซ้่ฏฏใ้ฟๆฎต่ฝไธๆททๅๅ่กจไปไผๅคฑ่ดฅใ
ๆจกๅๅฎไฝ
่พๅ ฅๆฏ ASR ่ฏๅซๅ็ไธญๆใ่ฑๆๆไธญ่ฑๆททๅๆๅญ๏ผ่พๅบๆฏๅฐฝ้ไฟ็ๅๆ็ๆด็ๆๆฌใ ๆจกๅไธ็ดๆฅๆฅๆถ้ณ้ข๏ผไธๆฏ่ฏญ้ณ่ฏๅซๆจกๅ๏ผไนไธๆฏ้็จ้ฎ็ญๅฉๆใ ๅพ ๆด็ๆๆฌไธญ็้ฎ้ขๅๅฝไปคๅฑไบๅฃ่ฟฐๅ ๅฎน๏ผไธๅบ่ขซๅ็ญๆๆง่กใ
ๆจกๅ่พๅบ็ดๆฅๅ ๅซๆ ็นใๆข่กๅๅ่กจ็ผๅท๏ผ่ฟไบๆ ผๅผไธๆฏๆจ็็คบไพไธญ็่งๅๅๅค็็ๆ็ใ ๆฌๆจกๅ็ๅฝๅไฝฟ็จ้ ็ฝฎไธๅฏ็จๅค้จๅ ๅฎนไฟๆคๆ ก้ช๏ผไธๅๅ ๆไธช่ฏๆๆฐๅญๆ ก้ชๅคฑ่ดฅ่้ๅๆดๆฎตๅๆใ ่ฟๆฏๅค้จๆๅก้ ็ฝฎ๏ผไธๆฏๆ้็ไธ้จๅใๆจ็็คบไพไน็ดๆฅ่ฟๅๆจกๅ็ๆ็ปๆใ
่ฎญ็ปๆฐๆฎไธๆนๆณ
ไธป่ฆไธญๆๆฅๆบไธบ CS2W ๅฃ่ฏญๅฐไนฆ้ข่ฏญๆ ๆณจๅฏน็็ญ้่ฎญ็ป้จๅใ ่พ ๅฉไฝฟ็จ Disfl-QA ่ฑๆๆนๅฃ่ฎญ็ปๆฐๆฎ๏ผไปฅๅ่ชๅถๅๆใไธไฟฎๆนใๅผ็จใ่ฏญๆฐใ็ฉบ่พๅบใๅ่กจๅๆฐๅญๅๆณๆ ทๆฌใ
ๆฌๆจกๅ็ป่ฟไธ้ถๆฎตๅ จ้ๆๆฌๅๆฐๅพฎ่ฐ๏ผไธๆฏ LoRA๏ผไนไธๆฏๅชไฟฎๆนๆ็คบ่ฏ๏ผ
| ้ถๆฎต | ๆฐๆฎไธ่ฎญ็ป |
|---|---|
| ็ฌฌไธ้ถๆฎต | 9645 ๆฌกๆ ทๆฌๅ็ฐ๏ผ1 epoch๏ผ603 ๆญฅ๏ผๅญฆไน ็ 1e-5 |
| ็ฌฌไบ้ถๆฎต | 14428 ๆกๅฎๅ จไธ้ฃๆ ผๅๆพๆฑ ๏ผๅญฆไน ็ 3e-6๏ผ้ๆฉ็ฌฌ 400 ๆญฅ checkpoint |
| ๆฌๆฌกๆดๆฐ | ไปไธไธๅๅธๆ้็ปง็ปญ่ฎญ็ป๏ผ7407 ๆกๅๆไธๅๆพๆ ทๆฌ๏ผ450 ๆญฅ๏ผๅญฆไน ็ 5e-6 |
ๆฌๆฌก 7407 ๆกๅ ๅซ 4907 ๆกๆฐๅข่ชๅถๅๆๆ ทๆฌๅ 2500 ๆกๆขๆ่ฎญ็ปๆฐๆฎๅๆพใ ๅ ๅฎน่ฆ็็ญๅไธพใ้ฟๅ่กจ้กนใไธญๆๅ้ฟๆไผฏๆฐๅญๅๆณใๆๆฏๅๅทใ็ๆฌใ้ช่ฏ็ ใๅคงๆดๆฐไธๆ่ฏญใ 7331 ไธชไธๅ็่พๅ ฅ/็ปๆ็ปๅ๏ผๆฒกๆไบ็ธๅฒ็ช็็ฎๆ ๆ ็ญพ๏ผๅ ๅซ 51 ๆก็ฉบ่พๅบ็็ฃใ ๆ ทๆฌๅ็ฐๆฐไธ็ญไบ็ฌ็ซไบบๅทฅๆ ๆณจๆฐใๅๆๆ ทๆฌไธ็ญไบ่ช็ถๆททๅ่ฏญ่จๅฃ่ฟฐๆฐๆฎใ
่ฎญ็ปไฝฟ็จ batch 4ใๆขฏๅบฆ็ดฏ็งฏ 4ใๆๅคงๅบๅ้ฟๅบฆ 512๏ผไป ๅฏน็ฎๆ ่พๅบ่ฎก็ฎๆๅคฑๅนถ็็ฃ EOSใ FP32 ไธปๅๆฐๅไผๅๅจใBF16 autocast๏ผๅๅธๆ้ไธบ BF16ใ็บฆ 7.52 ไบฟๅฏ่ฎญ็ปๆๆฌๅๆฐใ
ไธๆธธ็ๆฌ๏ผsuperwhisper/s1-mini@88f6b15896c73bbb13a3b596e0afe8ea0d5150b4ใ
ไธไธๅ
ฌๅผ็ๆฌ๏ผ782bc1216353eb4e31840292becbd525354e6435ใ
ๅ
ทไฝ่ฎญ็ปๅๆฐใๅๅธๅๆฐๆฎๆฅๆบ่ง training-provenance.json ไธ data-provenance.jsonใ
่ฎญ็ปๆชไฝฟ็จ็งๆๅฝ้ณๆ็จๆท่พๅ
ฅๅๅฒใๆฌไปๅบไธ้ๆฐๅๅธ่ฎญ็ปๆๆฌๆ็จๆทๆต่ฏๅๆใ
ๅฝๅๆ้็ๆต่ฏ็ปๆ
ๅไธ็ปๆขๆๅ ฌๅผๆฅๆบๅๅฝๆต่ฏ๏ผไธญๆๆด็ 200 ๆกใไธญๆไธไฟฎๆนๅ่ 100 ๆกใ่ฑๆ Disfl-QA 100 ๆกใ ไปฅไธๆฏๆช็ปๅค้จๅ ๅฎนๆ ก้ช็ๅๅงๆจกๅ่พๅบใๅๅไธคไธช checkpoint ๅ็จ BF16ใgreedyใbatch 16ใๆๅคง็ๆ้ฟๅบฆ 256๏ผ
| ๆๆ | ไธไธ็ step400 | ๅฝๅ format-numbers-v5 |
|---|---|---|
| ๅ จ้จ 400 ๆก่ฏๆณๅญ็ฌฆ้่ฏฏ็๏ผ่ถไฝ่ถๅฅฝ | 10.115% | 10.247% |
| ไธญๆๆด็๏ผ่ฏๆณๆดๅฅๅฎๅ จๅน้ | 83/200 | 81/200 |
| ไธญๆๆด็๏ผ่ฏๆณๅญ็ฌฆ้่ฏฏ็ | 16.790% | 16.712% |
| ไธญๆไธไฟฎๆนๅ่๏ผ่ฏๆณๆดๅฅๅฎๅ จไธ่ด | 92/100 | 91/100 |
| ไธญๆไธไฟฎๆนๅ่๏ผ่ฏๆณๅญ็ฌฆ้่ฏฏ็ | 1.440% | 1.592% |
| ่ฑๆๆด็๏ผ่ฏๆณๆดๅฅๅฎๅ จๅน้ | 64/100 | 60/100 |
| ่ฑๆๆด็๏ผ่ฏๆณๅญ็ฌฆ้่ฏฏ็ | 9.016% | 9.246% |
ๆฌๆฌกๆนๅไธป่ฆ้ๅฏนๅ่กจๅๆฐๅญๅๆณ๏ผๅนถ้ๆๆๆๆ ้ฝๆ้ซใ่ฑๆๅๆดไฝๅๅฝๅญๅจๅฐๅน ้ๆญฅใ ่ฏๆณๆๆ ๅ ้คๆ ็น๏ผไฟ็ๅคงๅฐๅใๆฐๅญใๅฆๅฎๅ่ฏๅบ๏ผ่กก้ไธๅ่ๆๆฌ็ไธ่ด็จๅบฆ๏ผไธ็ญไบไบบๅทฅ่ฏญไนๆญฃ็กฎ็ใ ่ฟไบๆต่ฏๅจๅผๅไธญๅคๆฌกไฝฟ็จ๏ผไธๆฏๆฐ็็ฌ็ซ็ฒๆต๏ผ่ฑๆๆนๅฃ้ไนไธ่ฝไปฃ่กจๆๆๆฅๅธธ่ฑๆ่พๅ ฅใ
ๅฆๆ 20 ๆกๆชๆพๅ ฅ่ฎญ็ป่พๅ ฅ็ๅฎๅ่ฏๆญ็จไพ๏ผ็ฆป็บฟ BF16 ๆน้่พๅบ 18/20 ไธๅ่ๅฎๅ จไธ่ด๏ผOrin vLLM ้ๆกๆจ็ 19/20 ๅฎๅ จไธ่ด๏ผ6 ๆกๅ่กจ็จไพๅ่พๅบ็ผๅท่กใ ่ฟ่กๆถ็ๅฉไฝๅทฎๅผๆฏไฟ็โไบไธชใๅ ญไธชโ๏ผ่้ๅ่็โ5ไธชใ6ไธชโใ่ฟ็ป็จไพๅ ๅซ้ฎ้ข่ฏๆญๅบๆฏ๏ผไธๆฏๅนฟๆณ่ฝๅๅบๅใ ไธๅๅ็ซฏๅ batch ็ BF16 greedy ่พๅบไนๅฏ่ฝไธๅใ ๅฎๆดๅฎไนไธ่ๅด่ง evaluation-summary.jsonใ
ไธไธ็็ 94.63% ๆๅธๅผบๅถ่ฑๆ token ๅ็กฎ็ๅฐๆชๅจๅฝๅๆ้ไธ้็ฎ๏ผไธ่ฝไฝไธบๆฌๆฌก็ๆฌๆ็ปฉใ ๅๅฒๅฎไนไธ่ฎกๆฐไฟๅญๅจ history/step400/token-metrics.jsonใ ไธๆธธๅ ฌๅผ็ 94.8% ็ผบๅฐ่ถณๅค็่ฏๅ็ป่๏ผไธ่ฝๅฃฐ็งฐ่ฟไบๆฌๅฐๆๆ ๆฏๅฎๆนๅๅฃๅพๅค็ฐใ
ๆผ่ฏป่ฑๆ็ๆจกๅๅ้ขๅค็
ๅบ็จๅฏๅ ๅๅนถๆผ่ฏปๅญๆฏ๏ผๅๆๆๆฌไบค็ปๆจกๅใ่ฟๆฏๅบ็จๅฑ่งๅ๏ผไธๅ ๅซๅจๆจกๅๆ้ไธญ๏ผไนไธๆฏ็ๆ็ปๆๅ็ๆฟๆขใ ้ๅธฆ็ input_preprocessing.py ๅชไพ่ต Python ๆ ๅๅบ๏ผPython 3.11+๏ผใ
ไฝฟ็จC-H-A-T-G-P-Tโไฝฟ็จchatgptA S T R AๆจกๅโAstraๆจกๅ- ๅๅนถๅๅ
่ฝฌๅฐๅ๏ผๅฅ้ฆๅชๅคงๅ้ฆๅญๆฏใๅญๆฏ้กบๅบไธๅ๏ผ
C-L-E-A-U-D-Eๅจๅฅไธญๅพๅฐcleaude๏ผๅฅ้ฆๅพๅฐCleaude๏ผไธไผๆ ่ชๆนๆClaudeใ - ้ๅธธๅๅนถไธไธชๅไปฅไธๅๅญๆฏ๏ผไธคไธชๅคงๅใ็ฉบๆ ผๅ้็ๅญๆฏไนๅฏๅๅนถ๏ผไพๅฆ
A Iใ - ไธ่ทจ่ถๆข่ก๏ผไธๆนๅไปฃ็ ๅญ้ข้ใ็ฝๅใ้ฎ็ฎฑๅ่ทฏๅพใไธคไธชๅญๆฏ็่ฟๅญ็ฌฆ่กจ่พพๅผ
X-Yไฟๆๅๆ ทใ
ๅ ไธ่ฝฝ่ๆฌๅฐๅบ็จ็ฎๅฝ๏ผ
hf download WIlfLin/S1-mini-Chinese input_preprocessing.py --local-dir .
ๅจๆ้ ๆจกๅๆถๆฏไนๅ่ฐ็จ๏ผ
from input_preprocessing import normalize_spelled_letters
transcript = "ๆไปๅคฉ่ฏ็จA S T R Aใ"
model_input, spelling_runs = normalize_spelled_letters(transcript)
# model_input == "ๆไปๅคฉ่ฏ็จastraใ"
# ๅฐ model_input ๆพๅจไธ้ขๆจ็็คบไพ็ user ๆงๅถ่กไนๅ๏ผๅ่ฐ็จๆจกๅใ
ๅบ็จ็ๅๆๆจกๅผๅบ่ทณ่ฟ่ฟไธๆญฅใ่ฟ็ปญๅญๆฏไนๅฏ่ฝ่กจ็คบ็ฌ็ซๅ้๏ผ่งๅๆ ๆณๅฎๅ จๅคๆญ่ฏญไน๏ผไปฃ็ ็ๆฎตไผ่ทณ่ฟ๏ผๅ ถไปๆญงไนไบค็ปๆจกๅใ ๅฅ้ฆๅคๆญ่ฆ็ๆๆฌๅผๅคดใๅธธ่งๅฅๆซๆ ็นไนๅใๆฐ่กๅๆฐ่กไธญ็ๅ่กจ้กนใๅทฒ็ป่ฟๅ็่ฑๆ่ฏไธๆนๅๅคงๅฐๅ๏ผๆจกๅไปๅฏ่ช่กๆด็่พๅ ฅใ ๆญคๅ 400 ๆกๆต่ฏ่กจๆฏๅๅงๆจกๅ่พๅบ๏ผไธๅ ๅซๆญค้ขๅค็ใๆ้ๆฒกๆๅ ๅขๅ ๆญค่ๆฌ่ๅๅใ
ๆจ็็คบไพ
้ช่ฏ็ฏๅขไธบ PyTorch 2.10ใTransformers 5.3ใไฝฟ็จ้ๅธฆ tokenizer/chat template๏ผๅ
ณ้ญ thinking๏ผ่พๅ
ฅๅธฆๆงๅถ่กใ
ไธ่ฌๆๆฌ้ๆฉ Structure: prose๏ผๆ็กฎๅไธพ้ๆฉ Structure: listsใ
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "WIlfLin/S1-mini-Chinese"
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"
"ไปๅคฉ็ไปปๅกๅฆไธ๏ผ็ฌฌไธๆฃๆฅๆๅกๆฅๅฟ๏ผ็ฌฌไบๅคไปฝๆฐๆฎๅบ๏ผ็ฌฌไธๆด็ไผ่ฎฎ่ฎฐๅฝใ"
)},
]
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():
result = model.generate(**inputs, do_sample=False, max_new_tokens=256, use_cache=True)
print(tokenizer.decode(result[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))
ๅชๆๅฃๆฐด่ฏๆถๅ ่ฎธ่พๅบ็ฉบๅญ็ฌฆไธฒใ่ฏทๆฑๅคฑ่ดฅใ่ถ ๆถๅ่พๅบๆชๆญๆฏๆๅกๅฑๅผๅธธ๏ผไธๅ ๅฎนไฟๆคๆ ก้ชไธๅใ ๆจกๅๅช็ๆๆๆฌ๏ผ่พๅ ฅๆณๆ่ฐ็จ็จๅบ่ด่ดฃไปฅๅฎๅ จ็ๆๆฌๆไบคๆนๅผไฟ็ๆข่ก๏ผ้ฟๅ ๆๆข่กๅฝไฝๅ้ๆ้ฎใ
้ๅถ
- ไบบๅใๆฐๅญใๅฆๅฎใไบๅฎๅๆ็กฎๆนๅฃไปๅฏ่ฝๆด็้่ฏฏใ
- ไธไฟ่ฏๆขๅคๅ้ณ้ๅญๆๅซๆทท่ฟ่ฏป๏ผๆจกๅไนๅฏ่ฝๆไป็ปๅฅๅ็ปๅฐพๅฝๅ ฅๅ่กจใ
- ๅคๆๅ่กจใ้ฟๆฎต่ฝๅ่ช็ถไธญ่ฑๆททๅ่พๅ ฅไป้ๆดๅคไบบๅทฅๆฐๆฎไธ้ชๆถใ
- ๅๆๆๆฌๆต่ฏไธ่ฝๆฟไปฃ็ๅฎ้บฆๅ ้ฃไธ็ฎๆ ๅบ็จ็่พๅ ฅๆต่ฏใ
- ๅ ณ้ญๅค้จๅ ๅฎนๆ ก้ชๅ๏ผๆจกๅๆฌ่บซ็้่ฏฏไผ็ดๆฅไฟ็ๅจ่พๅบไธญใ
่ฎธๅฏ่ฏไธๆฅๆบ็ฝฒๅ
ๆฌๆจกๅๆฏ S1-mini by Superwhisper ็ไฟฎๆน็ๆฌ๏ผๅบๅฑๆถๆๆฅๆบไธบ Qwen3-0.6Bใ
ไฟ็ๅๅง LICENSE ๅ NOTICEใๅ่ฎธๅฏ่ฏ่ฆๆฑ็ปง็ปญไฟ็ๅๆจกๅๅ็งฐไธไฝ่
็ฝฒๅ๏ผS1-mini-Chinese ๆฏๅพฎ่ฐ็ๆฌๅ็งฐ๏ผไธๆฟไปฃๅๅง็ฝฒๅใ
CS2W ๅฃฐๆ CC-BY-SA-4.0๏ผDisfl-QA ๅฃฐๆ CC-BY-4.0๏ผ่ชๅถๅๆๅ
ๅฎนไธบ CC0-1.0ใ
ๅ
ทไฝๆฅๆบๅ็ๆฌ่ง data-provenance.json๏ผไธๆๆฐๆฎ่ฎธๅฏ่ฏๆน็งฐไธบ็บฏ Apache-2.0ใ
ๅฝๅๆ้ SHA256๏ผ54602b3a2e0c484fe7ae796a4290969d3a1cc58e099c83e59948fe94889150a0ใ
- Downloads last month
- 104