max-0.1b-aimax-router

ํ•œ๊ตญ์–ด ์—…๋ฌด ์š”์ฒญ ํ•œ ์ค„(+์ฒจ๋ถ€ ์ข…๋ฅ˜)์„ ๋ฐ›์•„, ํ”„๋กฌํ”„ํŠธ์— ์ค€ ๊ธฐ๋Šฅ ๋ชฉ๋ก ์ค‘ ์–ด๋А ๊ธฐ๋Šฅ์˜ ์ผ์ธ์ง€์™€ ํ•„์š”ํ•œ ์—”์ง„ยท๊ธ‰ํ•จยทํ›„์† ์ž‘์—…์„ ์ž‘์€ JSON์œผ๋กœ ๋‚ด๋Š” 0.1B ํŒ๋‹จ ๋ชจ๋ธ(GGUF). AI MAX ์š”์ฒญ ํŒ๋‹จ์šฉ. ์ž์› ์šด์˜์˜ ํŒ๋‹จ ์žฌ๋ฃŒ์ผ ๋ฟ ์ตœ์ข… ๊ฒฐ์ •์€ ํ•˜์ง€ ์•Š๋Š”๋‹ค.

๋” ํฐ ํŒ: ์ƒˆ ๊ธฐ๋Šฅยท๋ถ€์ ์ ˆ ์š”์ฒญ์„ ๋” ์ž˜ ๊ณ ๋ฅด๋Š” 0.6B ํŒ quantumaikr/aimax-router-qwen3-0.6b (๊ฐ™์€ ํ”„๋กฌํ”„ํŠธ ํ˜•์‹; 0.1Bโ†’0.6B 2๋‹จ๋„ ์žฌ ๋ณด์•˜์œผ๋‚˜ 0.6B ๋‹จ๋… ๊ถŒ์žฅ โ€” ๋‹ค๋ฅธ ์ž‘์„ฑ์ž ๋ด‰์ธ ์„ธํŠธ์—์„œ ๋‹จ๋… 83.6% vs 2๋‹จ 80.3%, ์ด 0.1B ๋‹จ๋… 71.2%).

์ฃผ ํŒŒ์ผ(์ €์žฅ์†Œ ๋ฃจํŠธ)์ด ์ตœ์‹ ํŒ(v3)์ด๋‹ค. ์ด์ „ ์ฃผ ํŒŒ์ผ์€ v2b/์— ๊ทธ๋Œ€๋กœ ์žˆ๋‹ค(๋˜๋Œ๋ฆฌ๊ธฐ์šฉ, sha256 ๊ฐ™์Œ). preview/๋Š” ๋” ์ด์ „ ๋ฏธ๋ฆฌ๋ณด๊ธฐ์ด๋ฉฐ, preview/์˜ f16 GGUF๋Š” ๊ธด ๋ชฉ๋ก ํ”„๋กฌํ”„ํŠธ์—์„œ ์ถœ๋ ฅ์ด ๊นจ์ง€๋ฏ€๋กœ ์“ฐ์ง€ ๋ง ๊ฒƒ.

์ด ํŒ์˜ ์„ฑ๊ฒฉ (๋จผ์ € ์ฝ์„ ๊ฒƒ)

  • ๊ถŒ์žฅ ํ‘œ์ค€ ๋ชฉ๋ก(ํ˜„์žฌ ๊ธฐ๋Šฅ 10๊ฐœ)์— ๋งž์ถ˜ ํŒ์ด๋‹ค. ํ‘œ์ค€ ๋ชฉ๋ก์—์„œ ํ˜„์žฌ ๊ธฐ๋Šฅ ํŒ๋‹จ์€ ์ด์ „ ํŒ(v2b)๋ณด๋‹ค ์ข‹์•„์กŒ๋‹ค(์•„๋ž˜ ํ‘œ).
  • ๋ชฉ๋ก์— ์—†๋Š” ์ผ์„, ๋ชฉ๋ก์— ์žˆ๋Š” ๋น„์Šทํ•œ ๊ธฐ๋Šฅ์œผ๋กœ ๋ณด๋‚ด๋Š” ์ผ์ด v2b๋ณด๋‹ค ์žฆ๋‹ค(์ฒ˜์Œ ๋ณด๋Š” ๊ธฐ๋Šฅ ์‹œํ—˜์˜ "๋ชฉ๋ก์—์„œ ๋บŒ"์—์„œ none ์ •๋‹ต๋ฅ  68% โ†’ 48%). ๋น„์Šทํ•œ ๊ธฐ๋Šฅ์ด ๋งŽ์ด ์„ž์ธ ๋ชฉ๋ก์ผ์ˆ˜๋ก ์‹ฌํ•˜๋‹ค. router_post.v4_gate๊ฐ€ ์ด๋Ÿฐ ๊ฒฝ์šฐ์˜ ์ƒ๋‹น์ˆ˜๋ฅผ uncertain์œผ๋กœ ๋Œ๋ฆฐ๋‹ค(์•„๋ž˜ ํ‘œ).
  • ์ƒˆ ๊ธฐ๋Šฅ์„ ๋ชฉ๋ก์— ๋”ํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋Š” ์•„์ง ๊ฑฐ์˜ ๊ณ ๋ฅด์ง€ ๋ชปํ•œ๋‹ค(์ฒ˜์Œ ๋ณด๋Š” ๊ธฐ๋Šฅ 6์ข…: ๋ชฉ๋ก์— ๋„ฃ์–ด๋„ 3%). ์ƒˆ ๊ธฐ๋Šฅ์„ ์‹ค์ œ๋กœ ์“ฐ๋ ค๋ฉด ๊ทธ ๊ธฐ๋Šฅ ์š”์ฒญ ๋ฌธ์žฅ์„ ๋„ฃ์–ด ๋”ฐ๋กœ ๋‹ค์‹œ ํ•™์Šตํ•ด์•ผ ํ•œ๋‹ค.

ํŒŒ์ผ

ํŒŒ์ผ ๋ฌด์—‡
max-0.1b-aimax-router-Q4_K_M.gguf ๋ฐฐํฌ์šฉ(CPU ๊ฐ€๋Šฅ)
max-0.1b-aimax-router-bf16.gguf ๊ณ ์ •๋ฐ€(f16์€ ๊ธด ํ”„๋กฌํ”„ํŠธ์—์„œ ๊นจ์ ธ ์˜ฌ๋ฆฌ์ง€ ์•Š์Œ)
PROMPT.md ์ •ํ™•ํ•œ ํ”„๋กฌํ”„ํŠธยท๊ธฐ๋Šฅ ๋ชฉ๋ก ์ค„ ํ˜•์‹ยท๊ถŒ์žฅ ํ‘œ์ค€ ๋ชฉ๋กยท์ƒ์„ฑ ์„ค์ •ยท์ˆซ์ž ๋ณ‘ํ•ฉยท๊ฒŒ์ดํŠธ
standard_list.txt ๊ถŒ์žฅ ํ‘œ์ค€ ๋ชฉ๋ก(ํ˜„์žฌ ๊ธฐ๋Šฅ 10๊ฐœ + askยทnone), ๊ทธ๋Œ€๋กœ ์“ฐ๊ธฐ
router_post.py ๋ชฉ๋ก ๋ฌธ์ž์—ด, JSON ์Šคํ‚ค๋งˆ/GBNF, unitsยทseconds ๋ณ‘ํ•ฉ(v3_merge), ๊ฒŒ์ดํŠธ(v4_gate)
schema_example.json, grammar_example.gbnf ๊ถŒ์žฅ ํ‘œ์ค€ ๋ชฉ๋ก์šฉ ์ œ์•ฝ ์˜ˆ
scores.md ์ ์ˆ˜ํ‘œ(v2b์™€ ๋น„๊ต, ๊ฒŒ์ดํŠธ๋ณ„ ์ฑ„ํƒ๋ฅ ยท์ •ํ™•๋„)
TRAINING.md ํ•™์Šต ์„ค์ • ํ•œ ์ค„, ๋ฐ์ดํ„ฐ ์„ฑ๊ฒฉ
v2b/ ์ด์ „ ์ฃผ ํŒŒ์ผ(๋˜๋Œ๋ฆฌ๊ธฐ์šฉ)
v1/ ๊ธฐ๋Šฅ ๋ชฉ๋ก ์—†์ด ๊ณ ์ • 12๊ธฐ๋Šฅ๋งŒ ์“ฐ๋Š” ์ด์ „ ํ˜•์‹
SHA256SUMS

์ž…๋ ฅยท์ถœ๋ ฅ (v2b์™€ ๊ฐ™์Œ)

system ํ„ด: ๊ธฐ๋Šฅ ๋ชฉ๋ก: ๋‹ค์Œ ์ค„๋งˆ๋‹ค id|์„ค๋ช…|needs(needs๋Š” ,๋กœ, ์—†์œผ๋ฉด -), ๋์— ๊ณ ์ • ๋‘ ์ค„ ask|โ€ฆ|-, none|โ€ฆ|-. user ํ„ด: ์š”์ฒญ ํ•œ ์ค„ + \n[์ฒจ๋ถ€: โ€ฆ]. ์ถœ๋ ฅ: {"action","needs","urgency","follow_up","batchable"} 5๊ฐœ ํ‚ค. unitsยทseconds๋Š” ๊ทœ์น™์ด ์ฑ„์šฐ๊ณ , needs๋Š” ๋ชฉ๋ก ๊ฐ’์„ ์“ฐ๋ฉฐ, ๊ฒŒ์ดํŠธ๋ฅผ ๊ถŒ์žฅํ•œ๋‹ค. ์ž์„ธํ•œ ๊ฒƒ์€ PROMPT.md.

์ ์ˆ˜ ์š”์•ฝ (GPU, Q4_K_M, GBNF ์ œ์•ฝ; v2b โ†’ ์ด ํŒ)

์„ธํŠธ action unitsยทseconds ๊ฒŒ์ดํŠธ ์ฑ„ํƒ๋ฅ  / ์ฑ„ํƒ๋ถ„ action
gold 48 (ํ‘œ์ค€ ๋ชฉ๋ก) 97.9 โ†’ 97.9 92.3 โ†’ 92.3 91.7 / 97.7 โ†’ 87.5 / 100.0
๋ด‰์ธ 298 (ํ‘œ์ค€ ๋ชฉ๋ก) 90.3 โ†’ 93.6 78.8 โ†’ 87.5 64.8 / 99.0 โ†’ 66.1 / 99.5
๋ด‰์ธ 298 (์„ž์ธ ๋ชฉ๋ก) 82.9 โ†’ 88.9
์ฒ˜์Œ ๋ณด๋Š” ๊ธฐ๋Šฅ: ๋ชฉ๋ก์— ๋„ฃ์Œ 0 โ†’ 3
์ฒ˜์Œ ๋ณด๋Š” ๊ธฐ๋Šฅ: ๋ชฉ๋ก์—์„œ ๋บŒ(none) 68 โ†’ 48 v4_gate๊ฐ€ ์˜ค๋‹ต 52๊ฑด ์ค‘ 37๊ฑด์„ uncertain์œผ๋กœ ๋Œ๋ฆผ
JSON์€ ๋ชจ๋“  ์„ธํŠธ์—์„œ 100%. unitsยทseconds์˜ v3 ๊ฐ’์€ ์ƒˆ ์ˆซ์ž ๊ทœ์น™(v3_merge), v2b ๊ฐ’์€ ์ด์ „ ๊ทœ์น™ ๊ธฐ์ค€(๊ฐ™์€ v3 ์ถœ๋ ฅ์— ์ด์ „ ๊ทœ์น™์„ ์“ฐ๋ฉด ๋ด‰์ธ 82.5). ๊ฒŒ์ดํŠธ๋Š” v4_gate. ์ž์„ธํ•œ ํ‘œ๋Š” scores.md.

๊ทธ ๋ฐ–์˜ ํ•œ๊ณ„

  • gold 48์€ ๊ฐœ๋ฐœ ์ค‘ ํ‹€๋ฆฐ ์˜ˆ๋ฅผ ๋ณด๊ณ  ๋ชฉ๋ก ์„ค๋ช…์„ ๊ณ ์ณค์œผ๋ฏ€๋กœ ๋‚™๊ด€์ ์ด๋‹ค. ๋ด‰์ธ 298์ด ๋” ๊ณต์ •ํ•œ ์ง€ํ‘œ๋‹ค.
  • ํ•™์Šตยทํ‰๊ฐ€ ๋ฌธ์žฅ์€ ๋ชจ๋‘ ํ•ฉ์„ฑ์ด๋‹ค(ํšŒ์›์˜ ์‹ค์ œ ์š”์ฒญ ์—†์Œ). ์‹ค์ œ ์š”์ฒญ์—์„œ๋Š” ๊ทธ๋ฆผ์ž ์šด์˜์œผ๋กœ ํ™•์ธํ•  ๊ฒƒ.

ํ•™์Šต

๊ธฐ๋ฐ˜ ๋ชจ๋ธ quantumaikr/max-0.1b-v1 ๊ณ„์—ด SFT ์ฒดํฌํฌ์ธํŠธ์—์„œ ์ „์ฒด ๋ฏธ์„ธ์กฐ์ •(์ถœ๋ ฅ ํ† ํฐ์—๋งŒ ์†์‹ค). ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ํ•ฉ์„ฑ ๋ฌธ์žฅ + ์—ฌ๋Ÿฌ ๋งํˆฌ๋กœ ์“ด ๋ฌธ์žฅ, ์—…๋ฌด ๊ธฐ๋Šฅ ์•ฝ 680์ข…์˜ ์š”์ฒญ ๋ฌธ์žฅ(๊ธฐ๋Šฅ ๋ชฉ๋ก์„ ์ƒ˜ํ”Œ๋งˆ๋‹ค ๋ฐ”๊ฟˆ), ์ด์ „ ํŒ์ด ํ‹€๋ฆฌ๋˜ ์œ ํ˜•์„ ๋ณด๊ฐ•ํ•œ ๋ฌธ์žฅ. ํšŒ์›์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋Š” ์“ฐ์ง€ ์•Š์•˜๋‹ค. ๋ผ์ด์„ ์Šค Apache-2.0.


A 0.1B Korean request router (GGUF). The system turn carries a function list (id|description|needs lines, fixed ask/none lines at the end); the model outputs {"action","needs","urgency","follow_up","batchable"}. Tuned for the recommended standard list of 10 functions: sealed 298 requests action 93.6% (previous 90.3%). Weaknesses: out-of-list requests are more often sent to a neighbouring listed function (use router_post.v4_gate), and new functions added only to the list are rarely picked โ€” retraining is still needed. The previous main file is kept under v2b/. Base: quantumaikr/max-0.1b-v1. License: Apache-2.0. No member data used.

Downloads last month
-
GGUF
Model size
0.1B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for quantumaikr/max-0.1b-aimax-router

Quantized
(1)
this model