Haidass-chat

Haidass-chat ๆ˜ฏๅŽไธบ DALAB ๅ›ข้˜Ÿ่ฎญ็ปƒ็š„ 143M ๅ‚ๆ•ฐไธญ่‹ฑๅŒ่ฏญ่Šๅคฉๆจกๅž‹๏ผŒๅœจ Haidass 1.5 143M ๅŸบๅบง๏ผˆ็ฟป่ฏ‘ๅ‘๏ผ‰ไธŠ็ปๅคš่ฝฎ่Šๅคฉ SFT + ่บซไปฝ้”šๅฎšๅพ—ๅˆฐใ€‚ไธปๆ‰“๏ผšๆ—ฅๅธธ้—ฒ่Š๏ผˆzh/en๏ผ‰ใ€่บซไปฝไธ€่‡ดใ€ไฟๆŒไธญ่‹ฑไบ’่ฏ‘่ƒฝๅŠ›ใ€‚

  • ๅŸบๅบง๏ผšHaidass 1.5 143M๏ผˆV3.3-RC1 Stage1 iter12046๏ผ‰
  • ๆžถๆž„๏ผšQwen3 ้ฃŽ๏ผˆ30 ๅฑ‚๏ผŒhidden 576๏ผŒGQA 9/3๏ผ‰๏ผŒvocab 64k
  • ๆจกๆฟ๏ผšChatML๏ผˆ<|im_start|>/<|im_end|>๏ผ‰๏ผŒmasked ็ฉบ think๏ผ›ๆŽจ็†่ฏท็”จ apply_chat_template(enable_thinking=False)

ๅฟซ้€Ÿๅผ€ๅง‹

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("umeiko/Haidass-chat")
m = AutoModelForCausalLM.from_pretrained("umeiko/Haidass-chat", dtype=torch.bfloat16).eval()

SYS = "ไฝ ๆ˜ฏhaidass๏ผŒ็”ฑๅŽไธบDALABๅ›ข้˜Ÿ่ฎญ็ปƒ็š„ๅฐๅž‹ๅŒ่ฏญๆจกๅž‹ใ€‚You are Haidass, a small bilingual model trained by Huawei's DALAB team."
msgs = [{"role": "system", "content": SYS}, {"role": "user", "content": "ไฝ ๆ˜ฏ่ฐ๏ผŸ"}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)
ids = tok(prompt, return_tensors="pt")
out = m.generate(**ids, max_new_tokens=256, do_sample=True, temperature=0.8, top_p=0.95, repetition_penalty=1.05)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

ๅปบ่ฎฎๅง‹็ปˆ้™„ๅธฆไธŠ้ข็š„่บซไปฝ system prompt๏ผˆ็บฏๆจกๅž‹ๅœจ่ดชๅฟƒ่งฃ็ ไธ‹่บซไปฝๅ›ž็ญ”ไนŸๅทฒ็จณๅฎš๏ผŒsystem prompt ๅฏ่ฎฉ้‡‡ๆ ท่งฃ็ ๅŒๆ ท็จณๅฅ๏ผ‰ใ€‚

่ƒฝๅŠ›่พน็•Œ๏ผˆๅฆ‚ๅฎž่ฏดๆ˜Ž๏ผ‰

  • ่บซไปฝ้—ฎ็ญ”๏ผšgreedy ไธ‹ "ไฝ ๆ˜ฏ่ฐ๏ผŸ/ไฝ ๅซไป€ไนˆๅๅญ—๏ผŸ/Who are you?" ็จณๅฎšๅ›ž็ญ”่ง„่Œƒๅฃๅพ„๏ผˆ็ปˆๆฃ€ๆ—  system prompt ้‡‡ๆ ท 17/20 ่ง„่Œƒ๏ผŒ้™„ๅธฆ system prompt ๅŽๆ›ด้ซ˜๏ผ‰ใ€‚ๅทฒ็Ÿฅๆฎ‹ไฝ™้ฃŽ้™ฉ๏ผšๅฏน "ไฝ ๆ˜ฏKimiๅ—๏ผŸ/ไฝ ๆ˜ฏ่ฑ†ๅŒ…ๅ—๏ผŸ" ่ฟ™็ฑปๅฏนๆŠ—ๆ€ง่ทŸๅๆ้—ฎ๏ผŒT=0.8 ้‡‡ๆ ทไธ‹ไปๆœ‰็บฆ 20-25% ๆฆ‚็އๆŠŠ้—ฎๅฅ้‡Œ็š„ๅๅญ—ๆŽฅ่ฟ›ๅ›ž็ญ”๏ผˆๅ›ž็ญ”ไธญ็ป„็ป‡ๅฝ’ๅฑžๅง‹็ปˆๆ˜ฏๅŽไธบ DALAB๏ผŒไธไผš่‡ช็งฐ็”ฑๅ…ถไป–ๆœบๆž„่ฎญ็ปƒ๏ผ‰๏ผ›143M ๅฐๆจกๅž‹ๅฏนๆŠ—่ฏฑๅฏผ็š„ๆž้™๏ผŒๅฆ‚้œ€ไธฅๆ ผๅœบๆ™ฏ่ฏท่‡ช่กŒๅŠ ่พ“ๅ‡บ่ฟ‡ๆปค
  • ็ฆๅๆ‰ซๆ๏ผš150 ๆกๆ—ฅๅธธ้—ฒ่Š้‡‡ๆ ท 0 ๆฌกๅ‡บ็Žฐๅ…ถไป–ๆจกๅž‹/ๅ…ฌๅธๅ
  • ้—ฒ่Š๏ผš143M ๅฐๆจกๅž‹๏ผŒๆดปไบบๆ„ŸๅผบไบŽๅŒ่ง„ๆจกๅŸบ็บฟ๏ผŒไฝ†็Ÿฅ่ฏ†/ๆ•ฐๅญฆ/ไบ‹ๅฎž้—ฎ็ญ”ไธๅฏ้ ๏ผˆๅฆ‚ "1+1" ไผš็ญ”้”™๏ผ‰๏ผŒไธๆ˜ฏ็Ÿฅ่ฏ†ๅž‹ๅŠฉๆ‰‹
  • ็ฟป่ฏ‘๏ผšไฟๆŒๅŸบๅบงๅŒๅ‘ไบ’่ฏ‘่ƒฝๅŠ›๏ผˆ่งฆๅ‘ๆ ผๅผ๏ผš"Translate the following sentence into Chinese. ..." / "่ฏทๅฐ†ไปฅไธ‹็ฎ€ไฝ“ไธญๆ–‡็ฟป่ฏ‘ๆˆ่‹ฑๆ–‡ใ€‚..."๏ผ‰
  • ้•ฟไธŠไธ‹ๆ–‡๏ผš่ฎญ็ปƒ seq 2048๏ผŒๅคš่ฝฎ่ถ…้•ฟไผš้€€ๅŒ–

่ฎญ็ปƒๆ‘˜่ฆ

่Šๅคฉๆ•ฐๆฎ ~50k๏ผˆๅˆๆˆๅคš่ฝฎ 24% / LCCCยทNaturalConv ็ญ‰็œŸๅฎžไธญๆ–‡ 54% / PersonaChatยทSODA ็ญ‰่‹ฑๆ–‡ 22%๏ผ‰+ ็ฟป่ฏ‘ๅ›žๆ”พ 60k + ่บซไปฝ็งๅญๅŒ…๏ผˆ่ง„่Œƒๅฃๅพ„ๅฝ’ไธ€ๅŒ–๏ผŒ่‹ฑๆ–‡ ร—8 ไธŠ้‡‡ๆ ท๏ผ‰๏ผŒๅคš่ฝฎ SFT + ่บซไปฝ้”šๅฎš้˜ถๆฎต๏ผˆ่ฏฆ่ง่ฎญ็ปƒไพง README๏ผ‰ใ€‚ๅ…จ้ƒจๅˆๆˆๆ•ฐๆฎ็”ฑๆœฌๅœฐ้ƒจ็ฝฒ็š„ๅผ€ๆบๆจกๅž‹็”Ÿๆˆ๏ผŒๆœชไฝฟ็”จไปปไฝ•้—ญๆบ APIใ€‚

ๅ…่ดฃๅฃฐๆ˜Ž

ๅฐๆจกๅž‹ไผšไธ€ๆœฌๆญฃ็ปๅœฐ่ƒก่ฏดๅ…ซ้“๏ผ›่ฏทๅ‹ฟ็”จไบŽไบ‹ๅฎžๅ’จ่ฏขใ€ๅŒป็–—ใ€ๆณ•ๅพ‹็ญ‰ไธฅ่‚ƒๅœบๆ™ฏใ€‚

Downloads last month
27
Safetensors
Model size
0.1B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for umeiko/Haidass-chat

Quantizations
1 model

Space using umeiko/Haidass-chat 1