ai_smell_remover v1.1
ๆ AI ่ตท่็ไธญๆ้ๆฎตๆนๆ้ธญๅฅ๏ผyage.ai๏ผ็ๆ้ฃ๏ผ้ไฝ AI ๅณใ่ฟๆฏๅจ Qwen3.5-9B-Base ไธ่ฎญ็ป็ LoRA๏ผๅๅนถๅ้ๅๆ Q8_0 GGUFใ่ฎญ็ปๆนๆณๅๅ จ้จไปฃ็ ๅจ grapeot/voice-loraใ
ๅฎๅชๆนๆช่พๅๅฅๅญ๏ผไธๆนๆฎต่ฝ็ปๆใ่ฎบ่ฏ้กบๅบๅไบๅฎใ่พๅบไธๅฎ่ฆๅๅไธ้ไบๅฎๆผ็งปๆ ธๆฅ๏ผ่งไธๆโๅฑ้โใ
| ๆไปถ | ่ฏดๆ |
|---|---|
ai_smell_remover-v1.1-Q8_0.gguf |
ๆจกๅๆ้๏ผQ8_0๏ผ9.1 GB |
voice-lora-card.yaml |
็ป voice-lora rewrite ็จ็่ฐ็จๅก๏ผๆๅๆฌๆบ llama-server |
LICENSE |
Apache-2.0๏ผ็ปงๆฟ่ช Qwen3.5-9B-Base |
ๅฟซ้ๅผๅง
็ฎๅๅชๅจ Mac๏ผM3 Ultra๏ผllama.cpp ็ Metal ๅ็ซฏ๏ผไธๆต่ฏ่ฟใllama.cpp ไนๆฏๆ Linux ๅ Windows๏ผCPUใCUDA ็ญๅ็ซฏ๏ผ๏ผๆขๅนณๅฐๆถๅช้่ฆๆขๆๅฏนๅบๅนณๅฐ็ llama.cppใ
# 1. ไธ่ฝฝๆ้
hf download grapeot/ai_smell_remover ai_smell_remover-v1.1-Q8_0.gguf voice-lora-card.yaml --local-dir ./ai_smell_remover
# 2. ่ฃ
llama.cpp ๅนถ่ตทๆๅก๏ผMac๏ผbrew install llama.cpp๏ผLinux/Windows๏ผ็จ llama.cpp ็ release ๅ
๏ผ
llama-server -m ./ai_smell_remover/ai_smell_remover-v1.1-Q8_0.gguf --alias ai_smell_remover-v1.1 \
--port 8091 -c 16384 -np 4 -ngl 99 --jinja
# 3. ่ฃ
voice-lora๏ผๆนๅไธ็ฏ markdown ๆ็ซ
uv pip install git+https://github.com/grapeot/voice-lora
voice-lora rewrite article.md --card ./ai_smell_remover/voice-lora-card.yaml --out article.voice.md
voice-lora compare article.md article.voice.md # ้ๆฎตๅฏน็
ง้กต๏ผๅฎก้
็จ
voice-lora rewrite ไผ๏ผ
- ้ๆฎตๆนๅๆญฃๆๆฎต่ฝ๏ผๆ ้ขใๅผ็จๅใ่กจๆ ผใไปฃ็ ๅๅๅพ็ๅๆ ทไฟ็๏ผ
- ๆ้ๆๆฌๆพๅ้พๆฅ๏ผ
- ๆๆฎตๆนๅๅๆฐๅญๅฏนไธไธ๏ผๆ่ ้ฟๅบฆๆฏ่ถ ๅบ 0.6โ1.6๏ผๅฐฑ้่ฏไธๆฌก๏ผ่ฟไธ่กๅฐฑไฟ็ๅๆใ
ๅจ M3 Ultra ไธ๏ผ4 ่ทฏๅนถๅๆนๅไธ็ฏ 40 ๆฎต็ๆ็ซ ็บฆ 45 ็งใ
่ชๅทฑ่ฐ็จๆถๅฟ ้กปๆณจๆ
็จ voice-lora ๅฐฑไธ็จ็ฎก่ฟไธ่ใ่ชๅทฑๅๅฎขๆท็ซฏๆถ๏ผ
Prompt ๆ ผๅผ๏ผๅ่ฝฎ user ๆถๆฏ๏ผๅ
ณ้ญๆ่ๆจกๅผ๏ผchat template ๅๆฐ enable_thinking: false๏ผใ
ๆไธ้ข่ฟๆฎตๆๅญๆนๅๆ้ธญๅฅ็ๆ้ฃ๏ผๅ
ๅฎนๅ้กบๅบไธๅใ
ใไธๆ๏ผไป
ไพๅ่๏ผไธ่ฆๆนๅใ
<ไธไธๆฎต็ๅๆ๏ผ็ฌฌไธๆฎตๆฒกๆ่ฟไธ่>
ใ่ฆๆนๅ็ๆฎต่ฝใ
<่ฟไธๆฎต>
้ๆ ทๅๆฐ๏ผ
| ๅๆฐ | ๅผ | ๅๅ |
|---|---|---|
| temperature | 0.7 | greedy๏ผtemperature 0๏ผไธๆจกๅๅคงๅคๅจ็ งๆ่พๅ ฅ |
| top_p | 0.95 | |
| ้ๅคๆฉ็ฝ | 1.05 | llama.cpp ็จ repeat_penalty๏ผvLLM ็จ repetition_penalty |
| ๆฉ็ฝ็ชๅฃ | ่ฆ็ๆดไธช prompt | llama.cpp ่ฆไผ repeat_last_n: 4096๏ผ้ป่ฎคๅช็ๆ่ฟ 64 ไธช token๏ผ |
ๆฉ็ฝ็ชๅฃๆฏๆๅฎนๆ่ธฉ็ๅใ่ฆๆนๅ็ๅๆๅฐฑๅจ prompt ้๏ผ้ๅคๆฉ็ฝๅไฝไบ็ งๆๅๆ็ๆฆ็๏ผ่ฟๆฏๆจกๅ่ฏๆข่ฏดๆณ็ไธป่ฆๅๅ ใๅ 14 ๆฎตใๅ้ๆ ท 6 ๆฌก๏ผๅๅๆ็ๅนณๅ็ธไผผๅบฆ๏ผ
| ้จ็ฝฒ | ็ธไผผๅบฆ |
|---|---|
| vLLM๏ผๆฉ็ฝ้ป่ฎค่ฆ็ๅ จ้จ prompt๏ผ | 0.786 |
llama-server๏ผไผ repeat_last_n: 4096 |
0.784 |
| llama-server๏ผ้ป่ฎค 64 ไธช token ็ชๅฃ | 0.849 |
| ไธๅ ๆฉ็ฝ | 0.901 |
ไธ่ฆ็จ LM Studio๏ผๅฎไธ่ฎค repetition_penalty๏ผไนไธๆ repeat_last_n ไผ ็ปๅผๆ๏ผๆนๅไผๆๆพๅไฟๅฎใ
่ฎญ็ป
ๅบๅบง๏ผQwen3.5-9B-Baseใ
LoRA๏ผbf16๏ผrank 32ใalpha 32๏ผๅญฆไน ็ 1e-4๏ผไฝๅผฆ้็ซ๏ผ๏ผๆๆ batch 16ใๅ้ช่ฏ loss ๆไฝ็ checkpoint๏ผ็บฆ 0.19 epochใ
ๆฐๆฎ๏ผ้ธญๅฅๆฌไบบ็ 185 ็ฏไธญๆๅๅฎข๏ผ2,160 ไธชๆฎต่ฝ๏ผๆๆ็ซ ๅๅ่ฎญ็ปใ้ช่ฏใๆต่ฏ้ใๆฐๆฎ็ๆ้ ๆนๆณๅซๅๅๅๆ๏ผๅ ่ฎฉ AI ๆๅๆๆฎต่ฝๆนๅๆ AI ่ ๏ผๅๆๆฐๆฎๅฏนๅ่ฟๆฅ่ฎญ็ป๏ผAI ่ โ ๅๆ๏ผใๆนๅๆฅ่ชไธไธชๆจกๅๅฎถๆ๏ผ
- ไธไธชๅผๆบ 27B ๆจกๅ๏ผ5 ็ง prompt๏ผ
- Gemini 3.8 Flash๏ผ3 ็ง prompt๏ผ
- DeepSeek V4.1 Flash๏ผ3 ็ง promptใ
่ฎญ็ปๅฏนๅ ฑ 18,163 ๆกใๅๆๅๅๆๆฐๆฎ้ฝไธๅ ฌๅผใ
่ฏไผฐ
ๅๆฐๆฏโๅ้ธญๅฅๅ็โๅ็ฑปๅจๆฆ็ P(ไฝ่ )๏ผ็ปไธ็จ v1.1 ๆฐๆฎ่ฎญ็ป็ๅ็ฑปๅจๆๅใๅ็ฑปๅจ่ง ๆจกๅๅกใ
| ๆต่ฏ่พๅ ฅ | AI ่พๅ ฅ | v1 | ๆฌๆจกๅ๏ผv1.1๏ผ | ้ธญๅฅๅๆ |
|---|---|---|---|---|
| 27B ๆนๅ็็ๅบๆฎต่ฝ๏ผ291๏ผ | 0.09 | 0.77 | 0.76 | 0.75 |
| Gemini ๆนๅ็็ๅบๆฎต่ฝ๏ผ285๏ผ | 0.02 | 0.76 | 0.74 | 0.75 |
| DeepSeek ๆนๅ็็ๅบๆฎต่ฝ๏ผ289๏ผ | 0.25 | 0.74 | 0.76 | 0.75 |
| 4 ็ฏ AI ่ตท่็ๆด็ฏๆ็ซ ๏ผ147 ๆฎต๏ผ | 0.04 | 0.12 | 0.15 | โ |
- ็ๅบๆฎต่ฝไธ็ๅๆฐๅทฒ็ปๅๅๆๆๅนณ๏ผๅ็ฑปๅจๅจ่ฟ้ๅไธๅบ้ซไธใ
- ๆด็ฏ AI ๆ็ซ ็ๆๅๆฏ็ๅฎ็๏ผไฝๅน ๅบฆๆ้๏ผๅฎๅชๆขๆช่พ๏ผAI ๆ็ซ ็็ปๆๅ่ฎบ่ฏๅฅ่ทฏ่ฟๅจใ
- ไธไธ็ v1 ๅ่ฟไฝ่ ๆฌไบบ็็ฒ่ฏ๏ผ20 ๆฎตๆจกๅ่พๅบ้๏ผไฝ่ ๅค 16 ๆฎตๆฏ่ชๅทฑๅ็๏ผไฝ่ ่ชๅทฑ็็ๅๆ๏ผ่ฎคๅบไบ 17 ๆฎตใv1.1 ๆฒกๆๅ็ฌๅ็ฒ่ฏใ
- ๆฌ GGUF๏ผQ8_0๏ผๅ bf16 ๅ็ๅจ vLLM ไธ็ๅฏนๆฏ๏ผๅ ณๆ้ๆบๆงๆถ๏ผ40 ๆฎต้ 25 ๆฎต้ๅญ็ธๅ๏ผๆญฃๅธธ้ๆ ทๆถ๏ผๅๅๆ็็ธไผผๅบฆไธ่ด๏ผ0.784 ๅฏน 0.786๏ผใ
ๅฑ้
ไผๆไบๅฎๆผ็งป๏ผ่พๅบไธ่ฝ็ดๆฅๅฝๆ็ป็จฟใ ๆฐๅญไฟ้ฉๅชๆฆๅพไฝๆฐๅญๅ้ฟๅบฆใๅฎๆตๅบ็ฐ่ฟ็ๆผ็งป๏ผ
- ๅ ณ็ณปๆน้๏ผโ่ขซ X ๆถ่ดญ็ Yโๅๆโๅฏน X ๆๅ ฅโฆโฆๆถ่ดญ็ Yโ๏ผ
- ๅคๆญๅผบๅบฆ่ขซๆพๅคง๏ผโๆดๅคง็ๅๆฐโๅๆโๆๆฐ็บงโ๏ผ
- ๅ ๆๅๅฅ่ฎบ็นๆ้ๅฎๆกไปถ๏ผ
- ่ฏดๆๆๅๆ็ฌฌไธไบบ็งฐ๏ผ
- ่กๅ ไปฃ็ ็ๅๅผๅท่ขซๅปๆ๏ผ
- ่ฐ้ซๆธฉๅบฆๅ๏ผ็ผๅบๅๆๆฒกๆ็่ฑๆๅผ่ฏญใ
ๆนๅฎ่ฆ้ๆฎตๅฏน็ งๆนๅๅๅ๏ผๅชๆๆผ็งป็ๅญ่ฏๆนๅๅ็จฟ็่ฏดๆณ๏ผๆผ็งปๅคชๅคๅฐฑๆดๆฎตๆขๅๅๆใ
้ๆฎตๆนๅไผ่ฎฉๆ็ไธ็ปไธใ ๆน่ฟ็ๆฎต่ฝไผๅธฆไธไฝ่ ็ๆ็ไน ๆฏ๏ผ็ดๅผๅทใๅ่งๆฌๅทใไธญ่ฑๆไธ็็ฉบๆ ผ๏ผ๏ผๆฒกๆน็้จๅ่ฟๆฏๅ็จฟๅๆณใ
ไธๆน็ปๆ๏ผไธๆ ธๅฏนไบๅฎ๏ผๅ็จฟ้็้่ฏฏไผๅๆ ทไฟ็ใ
ๅช้ๅ AI ่ตท่็ๆๅญ๏ผ็จๅจไฝ่ ๆฌไบบๆๅ็ๆๅญไธๅชไผๆทปไนฑใ
่ฟๆฏไธไฝไฝ่ ็ๆ้ฃใ่ฏทไธ่ฆ็จๅฎๅๅ ไฝ่ ๆฌไบบๅ่กจๅ ๅฎนใ
English summary
ai_smell_remover rewrites AI-drafted Chinese, paragraph by paragraph, into the voice of the author ้ธญๅฅ (yage.ai). It changes wording only and keeps structure and facts. It is a LoRA on Qwen3.5-9B-Base (Apache-2.0), merged and quantized to Q8_0 GGUF.
- Training: reverse-synthesized pairs. The author's own paragraphs were rewritten into AI prose by an open 27B model, Gemini 3.8 Flash and DeepSeek V4.1 Flash, then the pairs were flipped. The data is not released.
- Serving:
llama-server. Tested only on macOS (M3 Ultra); llama.cpp also runs on Linux and Windows. - Sampling: temperature 0.7, top_p 0.95, repetition penalty 1.05 applied over the whole prompt (
repeat_last_n: 4096in llama.cpp). Do not use LM Studio, which drops that setting. - Always fact-check the output against the input and revert drifted phrases.
- Code and docs: https://github.com/grapeot/voice-lora.
- Downloads last month
- 659
8-bit
Model tree for grapeot/ai_smell_remover
Base model
Qwen/Qwen3.5-9B-Base