FinchAI-Rune2 v3

FinchAI ็ณปๅˆ—็ฌฌ 2 ไปฃ ยท v3 ็ปญ่ฎญ็‰ˆใ€‚ๅบ•ๅบง LFM3-1.7B๏ผŒ LoRA ๅพฎ่ฐƒ๏ผŒไธ“ๆณจๅทฅๅ…ท่ฐƒ็”จๅˆคๆ–ญๅŠ›ไธŽๅคš่ฝฎๅฏน่ฏ็จณๅฎšๆ€งใ€‚

็›ธๆฏ” v2๏ผˆwhcl412/FinchAI-Rune2-f16๏ผ‰๏ผš ่กฅไบ† 139 ๆกๅคš่ฝฎๆ ทๆœฌ๏ผŒๅคš่ฝฎๅค่ฏปไปŽ 1 ๆฌก้™ๅˆฐ 0 ๆฌก๏ผŒๅ…ถไฝ™่ƒฝๅŠ›ๅ…จ้ƒจไฟๆŒใ€‚

FinchAI ไธ‰ไปฃ่ƒฝๅŠ›ๅฏนๆฏ”

ๅ›พไธญๆ•ฐๆฎ๏ผšRune1 ไธบ 7 ้กนๆŽข้’ˆ ร— 5 ๆฌกๆŠฝๆ ท๏ผŒRune2 ไธบ 14 ้กนๆŽข้’ˆ ร— 2 ๆฌกๆŠฝๆ ท๏ผŒ ๅ„็ปดๅบฆๅ–่ฏฅ็ปดๅบฆไธ‹ๆ‰€ๆœ‰ๆŽข้’ˆ็š„ๅนณๅ‡้€š่ฟ‡็އใ€‚ไธคไปฃๅฃๅพ„ไธๅฎŒๅ…จไธ€่‡ด๏ผŒไป…ไพ›ๅ‚่€ƒ่ถ‹ๅŠฟใ€‚

ๆœฌไป“ๅบ“ -v3-f16๏ผˆbfloat16 ๅฎŒๆ•ดๆƒ้‡๏ผŒ3.24 GiB๏ผ‰
่ƒฝๅŠ›่Œƒๅ›ด ็บฏๆ–‡ๆœฌ๏ผˆไธๅซ่ง†่ง‰๏ผ‰
ไธŠไธ‹ๆ–‡็ช—ๅฃ 262144๏ผˆ256K๏ผ‰๏ผŒไธŽๅบ•ๅบงไธ€่‡ด๏ผ›ไฝ†ๅชๅœจ 1024 token ๅ†…่ฎญ็ปƒ่ฟ‡
ๅŠ ่ฝฝๆ–นๅผ AutoModelForCausalLM + trust_remote_code=True
่ฎธๅฏ ็ปงๆ‰ฟๅบ•ๅบง LFM Open License v1.0๏ผˆไธๆ˜ฏ Apache 2.0๏ผ‰

ไธ€ใ€v3 ๆ”นไบ†ไป€ไนˆ

v2 ๅฎžๆต‹ๅˆฐไธ€ไธชไธฅ้‡้—ฎ้ข˜๏ผš่ฟž็ปญ่Šๅˆฐ็ฌฌ 3~4 ่ฝฎๅผ€ๅง‹ๅค่ฏปโ€”โ€”ไธŠไธ‹ๆ–‡้‡Œ้‡ๅค็š„ๆจกๅผไผš่‡ชๆˆ‘ๅผบๅŒ–๏ผŒ ้—ฎไป€ไนˆ้ƒฝๅ›žๅŒไธ€ๅฅ๏ผˆๅฎžๆต‹ใ€Œ๐Ÿ˜Š ็ฌ‘่„ธใ€‚ใ€่ฟžไบ†ๅ››่ฝฎ๏ผ‰ใ€‚

v3 ่กฅไบ†ไธ“้—จๆ•ฐๆฎๆฅๆฒปๅฎƒ๏ผš

ๆ•ฐๆฎ็ฑปๅž‹ ๆกๆ•ฐ ๆ•™ไป€ไนˆ
short 61 ็”จๆˆทๅชๅ›žใ€Œๅ“ฆ/ๅ—ฏ/ๅฅฝ/๏ผŸใ€๏ผŒไธ€ไธคๅฅ่ฏ็กฎ่ฎคๅณๅฏ๏ผŒไธๅ†่งฃ้‡Šไธ€้
norepeat 37 ไธŠไธ€่ฝฎ็ญ”ไบ† A๏ผŒ่ฟ™่ฝฎ้—ฎๅˆซ็š„๏ผŒๅช็ญ”ๆ–ฐ็š„
carry 52 ไธŠไธ€่ฝฎ็ป™ไบ†ๅ…ทไฝ“ไฟกๆฏ๏ผŒ่ฟฝ้—ฎ่ฆๅผ•็”จๅฎƒ

ๅ…ณ้”ฎๅค„็†๏ผš็”Ÿๆˆ่ฟ™ไบ›ๆ•ฐๆฎๆ—ถ system ้‡Œๅ†™ๆ˜Žไบ†ใ€Œไธ่ฆ้‡ๅคใ€๏ผŒไฝ†่ฎญ็ปƒๆ—ถๆŠŠ system ๆขๅ›ž็ฎ€็Ÿญ็‰ˆโ€”โ€” ๅฆๅˆ™ๆจกๅž‹ๅชไผšใ€Œ็œ‹ๅˆฐๆ็คบๆ‰ไธ้‡ๅคใ€๏ผŒๆขไธชๅฎฟไธปๅฐฑๅคๅ‘ใ€‚

ไบŒใ€ๅฎžๆต‹ๅฏนๆฏ”

ๆŒ‰ 5 ไธช่ƒฝๅŠ›็ปดๅบฆๅฝ’ๅนถ๏ผˆๆฏไธช็ปดๅบฆๅ–่ฏฅ็ปดๅบฆไธ‹ๆ‰€ๆœ‰ๆŽข้’ˆ็š„ๅนณๅ‡้€š่ฟ‡็އ๏ผ‰๏ผš

่ƒฝๅŠ›็ปดๅบฆ Rune1 v2 Rune2 v2 Rune2 v3
่บซไปฝ่‡ช่ฟฐ 60% 100% 100%
ๅทฅๅ…ท่ฐƒ็”จ 67% 100% 100%
ๅฎ‰ๅ…จๆ‹’็ป 100% 83% 100%
็ผบๅ‚ๅ้—ฎ 60% 100% 100%
ๅคš่ฝฎๆ‰ฟๆŽฅ 60% 75% 100%
ๅคš่ฝฎๅค่ฏปๆฌกๆ•ฐ โ€” 1 0

โš ๏ธ ๅฃๅพ„่ฏดๆ˜Ž๏ผšRune1 ๆ˜ฏ 7 ้กนๆŽข้’ˆ ร— 5 ๆฌกๆŠฝๆ ท๏ผŒRune2 ๆ˜ฏ 14 ้กนๆŽข้’ˆ ร— 2 ๆฌกๆŠฝๆ ทใ€‚ ไธคไปฃๆต‹่ฏ•้›†ไธๅฎŒๅ…จ็›ธๅŒ๏ผŒไธŠ่กจๆ˜ฏๆŒ‰็ปดๅบฆๅฝ’ๅนถๅŽ็š„่ฟ‘ไผผๅฏนๆฏ”๏ผŒ็œ‹่ถ‹ๅŠฟ่€Œไธๆ˜ฏ็œ‹ๅฐๆ•ฐ็‚นใ€‚

v3 ็š„ๅฎž้™…ๅคš่ฝฎ่กจ็Žฐ๏ผš

่ฏดๅ‡ ไธชemoji โ†’ ๐Ÿ˜Š๐Ÿฅณ๐ŸŽ‰๐Ÿ‘๐Ÿ‘Ž๐Ÿ“ฆ๐Ÿ“„๐Ÿ“๐Ÿ–‹๏ธ๐Ÿ’ก๐Ÿ›๐Ÿ’ป๐Ÿš€๐Ÿ ๐ŸŒ™  ้œ€่ฆๆ›ดๅคšๆˆ–่€…ๆŒ‰ไธป้ข˜ๅˆ†็ป„๏ผŸ
ๅ•ฅๆ„ๆ€๏ผŸ    โ†’ ้‚ฃๆ˜ฏ็”จ่กจๆƒ…็ฌฆๅทไปฃๆ›ฟๆ–‡ๅญ—่ฏด่ฏ็š„ๆ„ๆ€ใ€‚๐Ÿ˜Š ็ฌ‘/ๅผ€ๅฟƒ ๐Ÿฅณ ๆƒŠๅ–œ/ๆฟ€ๅŠจโ€ฆ
ๅ“ฆ         โ†’ ๅฅฝ็š„๏ผŒ้šๆ—ถ้œ€่ฆใ€‚
ๆ‹œๆ‹œ       โ†’ ๅ—ฏใ€‚ไธ‹ๆฌกๆœ‰้œ€่ฆ้šๆ—ถๆฅใ€‚
๏ผŸ         โ†’ ไป€ไนˆๆ„ๆ€๏ผŸ่ฟ˜ๆ˜ฏ่ฏดไฝ ๆƒณ็ปง็ปญ่Š๏ผŸ

ไบ”่ฝฎๅ›ž็ญ”ๅ…จไธไธ€ๆ ทใ€‚

ไธ‰ใ€ๅฟซ้€Ÿๅผ€ๅง‹

pip install "transformers>=5.18.0" torch accelerate

ๅฟ…้กปๅธฆ trust_remote_code=True โ€”โ€” LFM3 ็š„ modeling ๆ–‡ไปถ่ทŸ็€ไป“ๅบ“ไธ€่ตทๅ‘ ๏ผˆmodeling_lfm3.py / configuration_lfm3.py / lfm3_conv.py๏ผ‰๏ผŒ ไธไผ ไผš็›ดๆŽฅๆŠฅ ModuleNotFoundErrorใ€‚

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "whcl412/FinchAI-Rune2-v3-f16"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype="bfloat16", device_map="auto", trust_remote_code=True)

tools = [{"type": "function", "function": {
    "name": "list_files", "description": "ๅˆ—ๅ‡บ็›ฎๅฝ•ๅ†…ๅฎน",
    "parameters": {"type": "object",
                   "properties": {"path": {"type": "string", "description": "็›ฎๅฝ•่ทฏๅพ„"}},
                   "required": ["path"]}}}]

msgs = [{"role": "user", "content": "็œ‹ไธ‹ /var/log ้‡Œๆœ‰ๅ“ชไบ›ๆ–‡ไปถใ€‚"}]
text = tok.apply_chat_template(msgs, tools=tools, tokenize=False,
                               add_generation_prompt=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device),
                     max_new_tokens=256, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))

ๅทฅๅ…ท่ฐƒ็”จๆ˜ฏ XML ๆ ผๅผ๏ผˆไธๆ˜ฏ JSON๏ผ‰๏ผš

<tool_call>
<function=list_files>
<parameter=path>
/var/log
</parameter>
</function>
</tool_call>

ๆ€่€ƒๅ†…ๅฎนๅœจ <think>โ€ฆ</think> ้‡Œ๏ผŒๅ’Œๆญฃๆ–‡ๅœจๅŒไธ€ๆก assistant ่พ“ๅ‡บไธญใ€‚

ๅ››ใ€ๅทฒ็Ÿฅ้—ฎ้ข˜

1. ๅคš่ฝฎไปๅปบ่ฎฎๅš้˜ฒๆŠค๏ผˆ่™ฝ็„ถๅทฒ็ปๅฅฝๅพˆๅคš๏ผ‰

v3 ๆŠŠๅค่ฏปๆฒปไฝไบ†๏ผŒไฝ†ไธŠไธ‹ๆ–‡็ช—ๅฃๅปบ่ฎฎไป็„ถๅˆซๅผ€ๅคช้•ฟ๏ผš

if len(msgs) > 9:                      # ๅช็•™ๆœ€่ฟ‘ 4 ่ฝฎ
    msgs = [msgs[0]] + msgs[-8:]

ไฟ้™ฉ่ตท่งๅŠ ไธชๅค่ฏปๆฃ€ๆต‹๏ผˆๆˆๆœฌๆžไฝŽ๏ผ‰๏ผš

if ans and last_ans and ans.strip() == last_ans:
    msgs = [msgs[0]]               # ็œŸๅ‡บ็Žฐๅค่ฏปไบ†ๅฐฑๆธ…็ฉบ้‡ๆฅ
    last_ans = None

2. ๅคš่ฝฎๆ‰ฟๆŽฅไพ่ต–ไธŠๆ–‡ๅ†™ๅฎž

ๅฆ‚ๆžœไธŠๆ–‡ๆ˜ฏใ€Œๆˆ‘ๆฒกๆŸฅ่ฟ‡ใ€่ฟ™็ง็ฉบ่ฏ๏ผŒ่ฟฝ้—ฎๆ—ถๅฎƒไผš็ปง็ปญ่ฟฝ้—ฎ่ทฏๅพ„ใ€‚ไธŠๆ–‡้‡Œ็ป™่ถณๅ…ทไฝ“ไฟกๆฏ ๏ผˆ"10:02ใ€10:05ใ€10:08"๏ผ‰๏ผŒๅฎƒๆ‰่ƒฝๆŽฅๅพ—ไฝใ€‚

3. ๅ้—ฎไผš้—ฎๅไธ€ๅฑ‚

้—ฎ"ๆŠŠ้‚ฃไธชๆ–‡ไปถๆ”นๅ"๏ผŒๅฏ่ƒฝๅ้—ฎ"ๅ“ชไธช็›ฎๅฝ•๏ผŸ"่€Œไธๆ˜ฏ"ๅ“ชไธชๆ–‡ไปถ๏ผŸ"ใ€‚

4. ๅทฅๅ…ท้กบๅบๆ•ๆ„Ÿ

ๅŒไธ€็ป„ๅทฅๅ…ทๅชๆข้กบๅบใ€้—ฎ้ข˜ไธ€ๅญ—ไธๆ”น๏ผŒ้€‰ๆ‹ฉไผšๅ˜ใ€‚ๆŠŠๆœ€ๅฏ่ƒฝ็”จๅˆฐ็š„ๅทฅๅ…ทๆŽ’ๅœจๅ‰้ข่ƒฝๆ˜Žๆ˜พๆ”นๅ–„ใ€‚

5. ็ป“ๆžœไธ่ถณๆ—ถไผš็ผ–ๆ•ฐๅญ—

่ฎฉๅทฅๅ…ท่ฟ”ๅ›ž็›ดๆŽฅๅฏๅพ—็ญ”ๆกˆใ€‚ๅˆซๆŒ‡ๆœ›ๅฎƒไปŽๅˆ—่กจ้‡Œๆ•ฐไธชๆ•ฐโ€”โ€”็ป™ count_files(pattern, path) ่ฟ™็งไธ“็”จๅทฅๅ…ทๆœ€็จณใ€‚

6. ้•ฟไธŠไธ‹ๆ–‡ๆœช้ชŒ่ฏ

็ช—ๅฃ 256K๏ผŒไฝ†่ฎญ็ปƒๆ ทๆœฌๆœ€้•ฟ 756 tokenใ€‚่ถ…่ฟ‡็บฆ 1K ไน‹ๅŽ็š„่กจ็Žฐๆฒกๆœ‰ไปปไฝ•ไฟ่ฏใ€‚

7. ๅฎ‰ๅ…จ่พน็•Œๆ˜ฏ"ๆ‹’็ป + ็ป™ๆ›ฟไปฃ"

่ฎฉๅฎƒๅˆ  /Users ๅ…จ้ƒจๆ–‡ไปถ๏ผŒๅฎƒไผšๆ‹’็ป๏ผŒไฝ†ๅฏ่ƒฝ็ป™ๅ‡บ find /Users -name "*.tmp" -delete ่ฟ™็ฑป้™ๅฎš็ฑปๅž‹็š„ๆ›ฟไปฃๅ‘ฝไปคโ€”โ€”ๆŒ‰่ฎพ่ฎก่ฟ™ๆ˜ฏๅ…่ฎธ็š„ใ€‚่ฆ็ปๅฏน็ฆๆญขไปปไฝ•ๅˆ ้™คๅ‘ฝไปค๏ผŒๅœจๅฎฟไธปไพงๅ†ๆ‹ฆไธ€ๅฑ‚ใ€‚

ไบ”ใ€ๆ€Žไนˆ่ฎญ็š„

  • ๅบ•ๅบง๏ผšLFM3-1.7B๏ผˆ1.73B๏ผŒไธ‰ๆททๅˆๆžถๆž„๏ผ‰
  • ๆ–นๆณ•๏ผšLoRA r=8 / alpha=16๏ผŒtarget ไธบ q/k/v/o_proj + in_proj + out_proj + gate_proj
  • v3 = ไปŽ v2 ็š„ adapter ็ปญ่ฎญ๏ผš+139 ๆกๅคš่ฝฎๆ ทๆœฌ๏ผŒ83 ๆญฅ๏ผˆ0.5 epoch๏ผ‰๏ผŒlr 5e-5๏ผŒ15 ๅˆ†้’Ÿ
  • ๆ•ฐๆฎๆ€ป้‡ 1339 ๆก๏ผŒๅ…ญ็ฑป๏ผšagent 28.5% / ๅคš่ฝฎ 21.7% / refuse 17.0% / code 14.9% / ้—ฒ่Š 9.6% / ่บซไปฝ 8.4%
  • ๆ•ฐๆฎๆฅๆบ๏ผšDeepSeek V4.1-Flash ่’ธ้ฆ๏ผˆๅ…ณ้”ฎ่ƒฝๅŠ›๏ผ‰+ ๆ‰‹ๅ†™่กฅๅ……

ๅ…ญใ€่ฎธๅฏ

็ปงๆ‰ฟๅบ•ๅบง LFM Open License v1.0๏ผŒไธๆ˜ฏ Apache 2.0ใ€‚ไป“ๅบ“้‡Œ้™„ไบ† LICENSE ๅŽŸๆ–‡ใ€‚

ๆกๆฌพ ๅ†…ๅฎน
ๅ•†็”จ้—จๆง› ๅนด่ฅๆ”ถๆœช่พพ 1000 ไธ‡็พŽๅ…ƒ๏ผˆ$10,000,000๏ผ‰ๆ—ถๅ…่ฎธๅ•†็”จ๏ผ›่พพๅˆฐๆˆ–่ถ…่ฟ‡ๅˆ™ไธๅœจๆœฌ่ฎธๅฏ่Œƒๅ›ดๅ†…๏ผŒ้œ€ไธŽ Liquid AI ๅฆ็ญพ
้ž่ฅๅˆฉไพ‹ๅค– ๅˆๆ ผ้ž่ฅๅˆฉ็ป„็ป‡ๅš้žๅ•†ไธš/็ ”็ฉถ็”จ้€”ไธๅ—้™
ๅ†ๅˆ†ๅ‘่ฆๆฑ‚ ๅฟ…้กป้™„่ฎธๅฏๅ‰ฏๆœฌ๏ผ›ไฟฎๆ”น่ฟ‡็š„ๆ–‡ไปถๆ˜พ่‘—ๆ ‡ๆณจ"ๅทฒไฟฎๆ”น"๏ผ›ไฟ็•™ๅŽŸๆœ‰็ฝฒๅ
ๅ•†ๆ ‡ ไธๅพ—ไฝฟ็”จ Liquid AI ๅ็งฐๆˆ–ๅ•†ๆ ‡๏ผˆๆ่ฟฐๆฅๆบ้™คๅค–๏ผ‰

ๆœฌๆจกๅž‹ๆ˜ฏๅœจ LFM3-1.7B ไธŠๅš LoRA ๅพฎ่ฐƒๅพ—ๅˆฐ็š„่ก็”Ÿไฝœๅ“๏ผŒๆกๆฌพๅŒๆ ท้€‚็”จใ€‚

ไธƒใ€้กน็›ฎไฟกๆฏ

Downloads last month
378
Safetensors
Model size
2B 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 whcl412/FinchAI-Rune2-v3-f16

Finetuned
(46)
this model