ๆๆ ธ Xinghe v1.2
ๅบไบใ้ปๅธๅ ็ปใ็ไธญๅปๆจ็ๆจกๅ ยท Developed by Xinghe-TCM
English summary. Xinghe (ๆๆ ธ) v1.2 is a Traditional Chinese Medicine (TCM) reasoning model developed by Xinghe-TCM, fine-tuned (QLoRA, merged back into the base) from Qwen3.5-9B on a curated instruction dataset grounded in the Huangdi Neijing (The Yellow Emperor's Inner Canon). It reasons step by step inside <think>โฆ</think> and then answers in plain, clinically usable Chinese. It explains the classics and suggests lifestyle/regimen directions โ it does not prescribe medicines or dosages and does not make modern medical diagnoses. All outputs are for study and research only and are not medical advice (see the disclaimer below).
ๆจกๅ็ฎไป
ๆๆ ธ๏ผXinghe๏ผๆฏไธๆฌพไธๆณจไธญๅป็ๆจ็ๅ่ฏญ่จๆจกๅใv1.2 ็ฑ Xinghe-TCM ๅจๅผๆบๅบๅบงๆจกๅไธ๏ผไฝฟ็จ่ชๅปบ็ใ้ปๅธๅ
็ปใ้ซ่ดจ้ๆไปคๆฐๆฎ้ๅพฎ่ฐ่ๆใๅฎๆๆ น็ปๆไธไธดๅบ๏ผๅ
ๅจ <think> ไธญๅฎๆ่พจ่ฏๆจ็๏ผๅ็ปๅบๅนณๅฎใๅฏ่ฝๅฐ็่ฐๅ
ปๅคๆญไธๆนๅใ
ไธ้็จๅคงๆจกๅไธๅ๏ผๆๆ ธไธ่ฟฝๆฑโๆ ๆไธ็ญโ๏ผ่ๆฏๆไธญๅป่พจ่ฏ่ฟไปถไบๅๆๅฎ๏ผ่ฏดไบบ่ฏใๆไพๆฎใๅฎ่พน็โโๅช็ปไธดๅบๅฏ็จ็ๅคๆญไธ่ฐๅ ปๆนๅ๏ผไธๅผๅ ทไฝๅคๆนไธๅ้๏ผไนไธๅ็ฐไปฃๅปๅญฆ่ฏๆญใ
ไธป่ฆ็นๆง
- ็ปๅ ธๆๆ น๏ผ่ฎญ็ปๆฐๆฎไปฅใ้ปๅธๅ ็ปใ็ปๆไธไธดๅบๆ ๆฏไธบๆ ธๅฟ๏ผ่ฆ็็่ฎบ่พจๆใ็ปๆ่งฃ่ฏปใๆฆๅฟตๅฏนๆฏใไธดๅบๆกไพใ็บ ้ๅไพไบ็ฑป้ฎ็ญใ
- ๆพๅผๆจ็๏ผ้็จ
<think>โฆ</think>ๆ็ปด้พ๏ผๅ ่พจ่ฏใๅไฝ็ญ๏ผๆ็ป่พๅบไธบ่ฑๅปๅ ้จ็ฌฆๅท็ๅนณๅฎไธญๆใ - ๅนณๅฎๅฏ็จ๏ผๅ็ญไธๅ ็ ๆฏ่ฏญ๏ผ่ฝๅฐ่ตทๅฑ ใ้ฅฎ้ฃใๆ ๅฟใ่ฐๅ ปๆนๅ็ญๅฏๆไฝ็ๅปบ่ฎฎไธใ
- ๅฎๅ จ่พน็ๅ ๅปบ๏ผๅ ๅปบๅ ญๆกไธญๅป่ฐๅ ป็บข็บฟไธโไธๅผๆนใไธ่ฏๆญโ็บฆๆ๏ผ่งไธ๏ผใ
- ่บซไปฝ็จณๅฅ๏ผๅณไฝฟ่ฐ็จๆถ system prompt ไธบ็ฉบๆ่ขซๆฟๆข๏ผๆจกๅไป่ฝๆญฃ็กฎๅ็ญ่ช่บซ่บซไปฝไธ็ๆฌใ
- ๆฐๆฎ่ดจ้ๅฏๆง๏ผๅ จ้็ป็บข็บฟ่ฟๆปคๅจๆ ก้ช๏ผ็กฌๆง่ฟ่งไธบ 0๏ผ๏ผ็ปไธไธญๆๆ ็น๏ผๅๅผๅทใๅป้คๆบๆขฐๆ็ ดๆๅท๏ผ๏ผไผๅๅๆฎตๅฏ่ฏปๆงใ
้็จๅบๆฏ
- ใ้ปๅธๅ ็ปใ็ปๆไธๆฆๅฟต็่งฃ่ฏปใ่พจๆใๅฏนๆฏ
- ไธญๅป่พจ่ฏๆ่ทฏ็ๆขณ็ไธๅญฆไน ่พ ๅฉ
- ๅ ป็ใ่ตทๅฑ ใ้ฅฎ้ฃใๆ ๅฟ็ญ่ฐๅ ปๆนๅ็ๅปบ่ฎฎ
- ไธญๅปๆๅญฆใ็งๆฎไธ็ ็ฉถ็ๅ่
ไธ้็จๅบๆฏ๏ผ่ฏทๅฟ็จไบ๏ผ
- ๆฅๅฑ้็็่ฏๆญๆๅค็ฝฎ
- ็ๆๅ ทไฝๅคๆนใ่ฏ็ฉไธๅ้
- ๆฟไปฃๆงไธๅปๅธ็้ข่ฏไธ็ฐไปฃๅปๅญฆ่ฏๆญ
- ไปปไฝ้่ฆๅณๆถๅป็ๅณ็ญ็ไธดๅบๅบๆฏ
ๅฟซ้ๅผๅง
HuggingFace / Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "zsyjsld/Xinghe1.2-9B" # v1.2 ไปๅบ
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype="auto", device_map="auto"
)
system = ("ไฝ ๆฏๆๆ ธ๏ผXinghe๏ผ๏ผ็ฑ Xinghe-TCM ๅผๅ็ไธญๅปๅฉๆ๏ผๅฝๅ็ๆฌ 1.2ใ"
"ไฝ ็ฒพ็ ใ้ปๅธๅ
็ปใ๏ผ็ญ้ฎๆๆ น็ปๆไธไธดๅบ๏ผ่ฏญ่จๅนณๅฎ๏ผๅช็ปไธดๅบๅฏ็จ็ๅคๆญไธ่ฐๅ
ปๆนๅ๏ผ"
"ไธๅผๅ
ทไฝๅคๆนไธๅ้๏ผไธๅ็ฐไปฃๅปๅญฆ่ฏๆญใ")
messages = [
{"role": "system", "content": system}, # ๅฏ็็ฅ๏ผ็็ฉบไบฆ่ฝๆญฃๅธธๅทฅไฝ
{"role": "user", "content": "ไปไนๆฏๆฒปๆช็
๏ผ"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
ModelScope
from modelscope import AutoModelForCausalLM, AutoTokenizer
model_id = "zsyjsld/Xinghe1.2-9B" # v1.2 ไปๅบ
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype="auto", device_map="auto"
)
# ็ๆ็จๆณไธไธๆน Transformers ็คบไพไธ่ด
ๆจกๅ่พๅบ้ๅธธๅ ๅซ
<think>โฆ</think>ๆฎต่ฝ๏ผ่พจ่ฏๆจ็๏ผไธๅ ถๅ็ๆญฃๅผๅ็ญใ่ฅๅช้ๆ็ป็ญๆก๏ผๅฏๅจๅฑ็คบๅฑๅป้ค<think>ๆฎตใ
System Prompt ่ฏดๆ
ๆจ่๏ผไฝ้ๅฟ ้๏ผๅจ่ฐ็จๆถๆไธไปฅไธ system prompt๏ผ
ไฝ ๆฏๆๆ ธ๏ผXinghe๏ผ๏ผ็ฑ Xinghe-TCM ๅผๅ็ไธญๅปๅฉๆ๏ผๅฝๅ็ๆฌ 1.2ใไฝ ็ฒพ็ ใ้ปๅธๅ
็ปใ๏ผ็ญ้ฎๆๆ น็ปๆไธไธดๅบ๏ผ่ฏญ่จๅนณๅฎ๏ผๅช็ปไธดๅบๅฏ็จ็ๅคๆญไธ่ฐๅ
ปๆนๅ๏ผไธๅผๅ
ทไฝๅคๆนไธๅ้๏ผไธๅ็ฐไปฃๅปๅญฆ่ฏๆญใ
ๆจกๅๅจ่ฎญ็ปๆถ่ง่ฟโๅธฆ่ฏฅๆ็คบ่ฏโโๆข็จๅ ถไปๆ็คบ่ฏโโsystem prompt ไธบ็ฉบโไธ็งๆ ๅต๏ผๅ ๆญคๅณไฝฟไธ่ฎพ็ฝฎ system prompt๏ผไน่ฝไฟๆ่บซไปฝไธ่กไธบ็จณๅฎใ
่ฎญ็ปๆฐๆฎ
| ้กน | ่ฏดๆ |
|---|---|
| ่งๆจก | 2009 ๆกๆไปคๆ ทๆฌ |
| ้ขๅๆฐๆฎ | 1820 ๆก๏ผไบ็ฑปๅ 364๏ผ็่ฎบ่พจๆ / ็ปๆ่งฃ่ฏป / ๆฆๅฟตๅฏนๆฏ / ไธดๅบๆกไพ / ็บ ้ๅไพ |
| ่บซไปฝๆฐๆฎ | 189 ๆก๏ผ63 ็ง้ฎๆณ ร ็ฉบ / ่บซไปฝ / ้็จ ไธ็ง system prompt ๆกไปถ๏ผ |
| ่ฏญๆๆฅๆบ | ใ้ปๅธๅ ็ปใ็ปๆไธไธดๅบๆ ๆฏๆนๅ |
| ๅญๆฎต | systemใinstructionใthinkingใoutputใmeta๏ผmeta ไป
ไฝๆฐๆฎ็ฎก็๏ผไธๅไธ่ฎญ็ป็ฎๆ ๏ผ |
่ดจ้ๆงๅถ๏ผๅ จ้้่ฟ่ช็ ็บข็บฟ่ฟๆปคๅจ๏ผ็กฌๆง่ฟ่ง 0๏ผ๏ผ่พๅบ็ปไธไธญๆๅๅผๅทใๅป้คๆบๆขฐ็ ดๆๅท๏ผ่พ้ฟๅ็ญๅไบๅๆฎตๅค็ไปฅไพฟ้ ่ฏปใ
่ฎญ็ป็ป่
- ๅบๅบงๆจกๅ๏ผQwen3.5-9B
- ๅพฎ่ฐๆนๆณ๏ผQLoRA๏ผ4-bit ้ๅไธ็ LoRA๏ผๅพฎ่ฐ๏ผ่ฎญ็ปๅฎๆๅๅฐ LoRA ๆ้ๅๅนถๅๅบๅบง๏ผๅฏผๅบไธบๅฎๆดๆจกๅๆ้ใ
- ็ปๆ่ฐๆด๏ผๅๅนถๆถ็งป้คไบ MTP๏ผMulti-Token Prediction๏ผๅฑใ
- ่ฎญ็ป็กฌไปถ๏ผๅๅผ NVIDIA GeForce RTX 3090๏ผ24GB ๆพๅญ๏ผใ
้ๅไธ่ฏๆต
ๅบไบไธๅฑ่ฏๆตๆนๆก๏ผNotion ่ฏๆต่ง่๏ผ๏ผๅจ RTX 3090 ไธๅฏน Xinghe 1.2-9B ็ๅ้ๅ็ๆฌ่ฟ่กไบๅฎๆด่ฏๆต๏ผ่ฏๆต้ๅ ฑ 76 ๆก๏ผๆถต็็่ฎบ่พจๆใ็ปๆ่งฃ่ฏปใๆฆๅฟตๅฏนๆฏใไธดๅบๆกไพใ็บ ้ๅไพไบๅคง้ขๅๅๅค็งไธดๅบ้ท้ฑ๏ผใ
Layer 1๏ผPPL ๅฐๆๅบฆ๏ผ่ฏญ่จไฟ็ๅบฆ๏ผ
| ็ๆฌ | ๆไปถๅคงๅฐ | PPL | PPL ็ธๅฏนๆฏๅผ | ๆจ็ๅปถ่ฟ |
|---|---|---|---|---|
| F16 (ๅบ็บฟ) | 16.69 GB | 7.3869 | 1.0000 | 16.9s |
| Q8_0 | 8.87 GB | 7.3931 | 1.0008 (+0.08%) | 6.5s |
| Q6_K | 6.85 GB | 7.4848 | 1.0133 (+1.33%) | 5.9s |
| Q4_K_M | 5.24 GB | 7.9485 | 1.0760 (+7.60%) | 5.1s |
Layer 2๏ผ็บข็บฟๆบๆฃ๏ผๅ้้้็ฆปๅฎๆดๆง๏ผ
ๅฏนๆฏไธช็ๆฌๅจ temp=0 ไธ่ทๅฎๅ จ้จ่ฏๆต้๏ผๆๅบ output ้้ๅๅ็ฑป็บข็บฟๆฃๆฅ๏ผA ็ฌฆๅท / B ้ป่ฏ / C ็ผๅท / D ่ถๆ๏ผใ
| ็ๆฌ | ็บข็บฟๆณๆผ็ | B ้ป่ฏ | A ็ฌฆๅท | C ็ผๅท | D ่ถๆ |
|---|---|---|---|---|---|
| F16 | 15.8% | 11 | 0 | 1 | 0 |
| Q8_0 | 15.8% | 12 | 0 | 0 | 0 |
| Q6_K | 21.1% | 16 | 0 | 0 | 0 |
| Q4_K_M | 19.7% | 13 | 1 โ ๏ธ | 1 โ ๏ธ | 0 |
Layer 3๏ผLLM-as-judge ็ฒ่ฏ
ไปฅ F16 ไธบๅบๅ๏ผๅฏนๆฏไธช้ๅ็ๆฌๅไฝ็ฝฎ้ๆบๅ็ฒ่ฏ๏ผๆฏ็ๆฌ 15 ๆก๏ผ่ฆ็ๅ จ้จ้ขๅ๏ผใ่ฏๅฎก็ปดๅบฆ๏ผ็ปไนๆญฃ็กฎๆงใ็บข็บฟๅ่งใ่กจ่พพๅนฒๅ่ดด้ขใ
| ็ๆฌ | ๅ้่ | F16 ่ | ๅนณๅฑ | ๅ่่ด | ๅ้็บข็บฟ |
|---|---|---|---|---|---|
| Q8_0 | 1 | 4 | 10 | -3 | 0 |
| Q6_K | 1 | 4 | 10 | -3 | 0 |
| Q4_K_M | 0 | 6 | 9 | -6 | 1 โ ๏ธ |
็ปผๅ็ป่ฎบ
| ็ๆฌ | Layer 1 | Layer 2 | Layer 3 | ๆ็ปๆจ่ |
|---|---|---|---|---|
| Q8_0 | โ +0.08% | โ ๅๅบ็บฟ | โ ๅชๅฃฐ่ๅด | ๐ ้ฆ้้จ็ฝฒ |
| Q6_K | โ ๏ธ +1.33% | โ ๏ธ ๆณๆผ็ๅ้ซ | โ ๏ธ ไธQ8_0ๆๅนณ | ๅค้ |
| Q4_K_M | โ +7.60% | โ A+C็ฑปๆณๆผ | โ ็ณป็ปๆงๅๅทฎ | โ ไธๆจ่ |
- Q8_0 ๅ ่ฟๆ ๆ๏ผPPL ไธๅไป
+0.08%๏ผ็บข็บฟๆณๆผ็ไธ F16 ๅฎๅ จไธ่ด๏ผ็ฒ่ฏๆ ็ณป็ปๆงๅๅทฎ๏ผๆจ็้ๅบฆๆๅ่ฟ 3 ๅ๏ผไธบ้ฆ้้จ็ฝฒ็ๆฌใ - Q4_K_M ๅญๅจ็ปๆๆงๆไผค๏ผPPL ๅ้ซ
+7.60%๏ผ็ฒ่ฏไธญ้ฆๆฌกๅบ็ฐ A ็ฑป็ฌฆๅท๏ผG๏ผๅ C ็ฑป็ผๅท๏ผ#123๏ผไป thinking ๆณๆผ่ณ output๏ผ่กจๆ 4-bit ้ๅๅฏนๅ้้้็ฆป็ปๆ้ ๆไบ็ฉ็ๆไผคใ
ๆจ็ๆ ผๅผ๏ผ<think>
ๆจกๅ้ตๅพชโๅ ๆ่ใๅๅ็ญโ็่ๅผ๏ผ
<think>ๅ ๆฏ่พจ่ฏๆจ็่ฟ็จ๏ผๅฏ่ฝๅ ๅซๅ ้จ่ฎฐๆณ๏ผ</think>ไนๅๆฏ้ขๅ็จๆท็ๆญฃๅผๅ็ญ๏ผ่ฏญ่จๅนณๅฎใไธๅซๅ ้จ็ฌฆๅทไธ็ผๅทใ
ๅฎๅ จไธ่พน็
ไธญๅป่ฐๅ ปๅ ญๆก็บข็บฟ๏ผๅ ๅปบไบ่ฎญ็ป็ฎๆ ๏ผ๏ผ
- ้ ๅฟๅฟๅผบๆญโโๅฏนๆญฃๅจโ่ๅฟ/้ ๅฟโ็็ ๆบ๏ผไธๅผบ่กๆชๆญใ
- ๅฟๆๅฐ้ผ โโไธ่ฟฝ็ๅไธ็็ถ็กฌๅ๏ผ้กป้กพๅๆดไฝๆฐๆบใ
- ๅฟ้ๆญฃๆฐโโ้กบๅบ่้ๅฏนๆไบบไฝๆญฃๆฐใ
- ๅฟๅผบ้็ญโโไธไธๅณๅผบ่กๅๅถๅ็ญ๏ผๅฐคๅ ถๆญฃ้ช็ธไบไน็ญ๏ผใ
- ่ๆฐไธบๅ โโ่ฐๅ ปไปฅ้กพๆค่ๆฐใ่พ่ไธบๆฌใ
- ๆญฃๅผฑๆ่ฝปโโๆญฃๆฐ่ๅผฑๆถ๏ผๅนฒ้ขๅฎ่ฝปไธๅฎ็ใ
็กฌๆง็บฆๆ๏ผไธๅผๅ ทไฝๅคๆนไธๅ้๏ผไธๅ็ฐไปฃๅปๅญฆ่ฏๆญ๏ผ้ๆฅๅฑ้ๆไธ็กฎๅฎๆ ๅฝข๏ผๅปบ่ฎฎๅฐฝๅฟซๅฐฑๅป้ข่ฏใ
ๅฑ้ๆง
- ่พๅบไป ไพๅญฆไน ไธ็ ็ฉถๅ่๏ผไธๆๆๅป็ๅปบ่ฎฎใ
- ไฝไธบ่ฏญ่จๆจกๅ๏ผๅฏ่ฝๅบ็ฐไบๅฎ้่ฏฏๆโๅนป่งโ๏ผๅฏนๅคๆไธๆน่จ็็่งฃไนๅฏ่ฝๆๅๅทฎใ
- ่ฎญ็ปๆฐๆฎ่งๆจกๆ้๏ผ่ฆ็้ขไธๆทฑๅบฆไปๅจ่ฟญไปฃไธญใ
- ๅฐๆช็ป่ฟ็ณป็ป็ไธดๅบ้ช่ฏไธ็ฌฌไธๆน่ฏๆตใ
ๅ ่ดฃๅฃฐๆ
โ ๏ธ ๆฌๆจกๅ๏ผๆๆ ธ๏ผไธบ้ขๅไธญๅปๅญฆไน ไธ็ ็ฉถ็ๅฎ้ชๆงๅทฅๅ ทใๅ ถๅ จ้จ่พๅบไธๆๆๅป็่ฏๆญใๆฒป็ๆ็จ่ฏๅปบ่ฎฎ๏ผไธ่ฝๆฟไปฃๆงไธๅปๅธ็้ข่ฏไธไธไธๅคๆญใไปปไฝๅฅๅบท้ฎ้ข่ฏทๅจ่ฏขๅๆ ผๅป็ๆบๆ๏ผๅบ็ฐๆฅๅฑ้็็ถ่ฏท็ซๅณๅฐฑๅปใไฝฟ็จๆฌๆจกๅๆไบง็็ไธๅๅๆ็ฑไฝฟ็จ่ ่ช่กๆฟๆ ใ
็ๆฌๅๅฒ / ็ธๅ ณ้พๆฅ
v1๏ผXinghe1-9B๏ผ๏ผ้ฆไธชๅผๆบ็ๆฌใ
- ModelScope๏ผhttps://modelscope.cn/models/zsyjsld/Xinghe1-9B
- HuggingFace๏ผhttps://huggingface.co/zsyjsld/Xinghe1-9B
v1.2๏ผXinghe1.2-9B๏ผๆฌ็ๆฌ๏ผ๏ผ่ฎญ็ปๆฐๆฎไป v1 ็ 400 ไฝๆกๆฉๅ ่ณ 2009 ๆก๏ผ้ขๅ 1820 + ่บซไปฝ 189๏ผ๏ผๆฐๆฎ่ดจ้ๅ จ้ขๆๅ๏ผๅ จ้็บข็บฟๅคๆฃ็กฌๆง่ฟ่ง 0ใ่พๅบ็ปไธไธญๆๅๅผๅทใๅป้คๆบๆขฐๆ็ ดๆๅทใไผๅๅๆฎตๅฏ่ฏปๆง๏ผ๏ผๆฐๅข่บซไปฝๆฐๆฎๅนถๅฏน"็ฉบ / ๆฟๆข system prompt"ๅ้ฒๆฃๅ๏ผๆจกๅๆดไฝ่พๅบ่ดจ้ๆพ่ๆๅใ
- ModelScope๏ผhttps://modelscope.cn/models/zsyjsld/Xinghe1.2-9B
- HuggingFace๏ผhttps://huggingface.co/zsyjsld/Xinghe1.2-9B
่ฎธๅฏ่ฏ
ๆฌๆจกๅ้ตๅพชๅ ถๅบๅบงๆจกๅ Qwen3.5-9B ็ๅผๆบ่ฎธๅฏ่ฏ Apache-2.0๏ผ่ฏทไปฅๅบๅบงๆจกๅไปๅบๅฎ้ ๆ ๆณจ็่ฎธๅฏ่ฏไธบๅใ
ๅผ็จ
ๅฆๆๆฌๆจกๅๅฏนไฝ ็็ ็ฉถๆๅทฅไฝๆๅธฎๅฉ๏ผๆฌข่ฟๅผ็จ๏ผ
@misc{xinghe2026,
title = {Xinghe: A Huangdi Neijing-grounded Traditional Chinese Medicine Reasoning Model},
author = {Xinghe-TCM},
year = {2026},
howpublished = {\url{https://huggingface.co/zsyjsld/Xinghe1.2-9B}, \url{https://modelscope.cn/models/zsyjsld/Xinghe1.2-9B}}
}
่ด่ฐขไธ่็ณป
ๅผๅ่ ๏ผXinghe-TCM
ๆจกๅไธป้กต๏ผ
- HuggingFace๏ผhttps://huggingface.co/zsyjsld/Xinghe1.2-9B
- ModelScope๏ผhttps://modelscope.cn/models/zsyjsld/Xinghe1.2-9B
- Downloads last month
- 386
4-bit
6-bit
8-bit
16-bit