thaki-AI commited on
Commit
21f37c3
ยท
verified ยท
1 Parent(s): 52c1552

Card: separate E0, K2-only and routed columns; note 6/200 demo names routed to E0

Browse files
Files changed (2) hide show
  1. MODEL_CARD.ko.md +7 -2
  2. README.md +7 -2
MODEL_CARD.ko.md CHANGED
@@ -43,7 +43,9 @@ top10 = np.argsort(-(q @ image_vectors.T), axis=1)[:, :10]
43
 
44
  **๋ชจ๋ธ ๋‹จ๋… R@10**
45
 
46
- | ์งˆ์˜ ์„ธํŠธ | n | E0 | ์ด์ „ K1 | **K2** | ๊ธฐ๋ณธ ๋ผ์šฐํŒ… | ๊ต์‚ฌ(1.5 GB) |
 
 
47
  |---|---|---|---|---|---|---|
48
  | Q-TITLE-ko: ํ•œ๊ตญ์–ด ์ƒํ’ˆ๋ช…(ABO ํ•œ๊ตญ์–ด ๋ฆฌ์ŠคํŒ…) | 889 | 0.368 | 0.231 | **0.645** | 0.643 | 0.682 |
49
  | kobench: ABO ์†์„ฑ ๊ธฐ๋ฐ˜ ํ•œ๊ตญ์–ด ํ…œํ”Œ๋ฆฟ ์งˆ์˜ | 173 | 0.104 | 0.069 | **0.318** | 0.318 | 0.491 |
@@ -54,11 +56,12 @@ top10 = np.argsort(-(q @ image_vectors.T), axis=1)[:, :10]
54
  - **ํ•œ๊ตญ์–ด ์ƒํ’ˆ๋ช…์—์„œ K2 ๋Š” E0 ๋ณด๋‹ค R@10 ์ด 0.277 ๋†’์Šต๋‹ˆ๋‹ค** [+0.243, +0.309](์ƒํ’ˆ ์œ ํ˜• ๊ตฐ์ง‘ 150๊ฐœ). ์ด๋ฒˆ ๊ณต๊ฐœ์˜ ๊ทผ๊ฑฐ๊ฐ€ ์ด ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.
55
  - kobench ์—์„œ๋„ K2 โˆ’ E0 ๋Š” +0.214 [+0.113, +0.315] ์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ ์ƒํ’ˆ ์œ ํ˜• ๊ตฐ์ง‘์ด 24๊ฐœ๋ฟ์ด๋ผ ํŒ์ • ๊ธฐ์ค€(30๊ฐœ)์— ๋ชป ๋ฏธ์ณ "๋ถˆํ™•์‹ค"๋กœ ๋‘ก๋‹ˆ๋‹ค.
56
  - K-Fashion(ํ•œ๊ตญ์–ด ํŒจ์…˜ ์†์„ฑ, ํ•™์Šต ๋ถ„ํฌ ์•ˆ, P@1, n = 626)์—์„œ๋Š” E0 0.005, ์ด์ „ K1 **0.403**, K2 0.289 ์ž…๋‹ˆ๋‹ค. **์ด ์„ธํŠธ์—์„œ๋Š” K2 ๊ฐ€ ์ด์ „ ํ•œ๊ตญ์–ด ๋ชจ๋ธ๋ณด๋‹ค ๋‚ฎ์Šต๋‹ˆ๋‹ค.**
 
57
  - ์˜์–ด ์ƒํ’ˆ๋ช…์€ ๊ทธ๋Œ€๋กœ์ด๊ณ (โˆ’0.001 [โˆ’0.009, +0.008]), ๋ฆฌ์ŠคํŒ… ํ‚ค์›Œ๋“œ๋Š” ์กฐ๊ธˆ ๋‚ฎ์•„์กŒ์Šต๋‹ˆ๋‹ค(โˆ’0.018 [โˆ’0.030, โˆ’0.006]). ์‹ค์ œ ์‡ผํ•‘ ๊ฒ€์ƒ‰์–ด๋Š” ์˜คํžˆ๋ ค ์˜ฌ๋ž์Šต๋‹ˆ๋‹ค(+0.100 [+0.059, +0.140]). ๊ธฐ๋ณธ ๋ผ์šฐํŒ…์€ ์˜์–ด๋ฅผ E0 ๋กœ ๋ณด๋‚ด๋ฏ€๋กœ ์˜์–ด ์ˆ˜์น˜๋Š” E0 ์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์˜์–ด ํŠธ๋ž˜ํ”ฝ์ด ESCI ์™€ ๋น„์Šทํ•˜๋‹ค๋ฉด `route="ko"` ๋ฅผ ๊ฒ€ํ† ํ•ด ๋ณด์„ธ์š”.
58
 
59
  **BM25 ์™€ ๊ฒฐํ•ฉ**(๋‘ ์ƒ์œ„ 10๊ฐœ ๋ชฉ๋ก์„ RRF, k = 60 ์œผ๋กœ ํ•ฉ์นจ. ํ•œ๊ตญ์–ด ์งˆ์˜์˜ BM25 ์ƒ‰์ธ์€ ์ƒํ’ˆ์˜ ํ•œ๊ตญ์–ดยท์˜์–ด ์ด๋ฆ„)
60
 
61
- | ์งˆ์˜ ์„ธํŠธ | BM25 R@1 | BM25 R@10 | BM25 + K2 R@1 | BM25 + K2 R@10 |
62
  |---|---|---|---|---|
63
  | Q-TITLE-ko | 0.911 | 0.993 | 0.703 | 0.991 |
64
  | kobench | 0.139 | 0.405 | 0.179 | 0.468 |
@@ -69,6 +72,8 @@ top10 = np.argsort(-(q @ image_vectors.T), axis=1)[:, :10]
69
  - ESCI(์˜์–ด)์—์„œ๋Š” K2 ๋ฅผ ๊ฒฐํ•ฉํ•˜๋ฉด BM25 ๋ณด๋‹ค R@10 ์ด +0.141 [+0.105, +0.181] ์˜ค๋ฆ…๋‹ˆ๋‹ค.
70
  - ์ด ๋ชจ๋ธ์˜ ์“ฐ์ž„์ƒˆ๋Š” ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค. ์ƒํ’ˆ ๋ชฉ๋ก์— ๋งž๋Š” ํ•œ๊ตญ์–ด ํ…์ŠคํŠธ๊ฐ€ ์—†์„ ๋•Œ ํ•œ๊ตญ์–ด ๊ฒ€์ƒ‰์–ด๋กœ **์ƒํ’ˆ ์ด๋ฏธ์ง€**๋ฅผ ์ฐพ๋Š” ๊ฒƒ, ๊ทธ๋ฆฌ๊ณ  ํ‚ค์›Œ๋“œ ๊ฒ€์ƒ‰์˜ ์ƒ์œ„ 10๊ฐœ ํ›„๋ณด๋ฅผ ๋„“ํžˆ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฒซ ๊ฒฐ๊ณผ๋ฅผ ์ •ํ•˜๋Š” ์ˆœ์œ„ ๋ชจ๋ธ๋กœ ์“ฐ์ง€ ๋งˆ์„ธ์š”.
71
 
 
 
72
  ์งˆ์˜ ๊ธธ์ด: Q-TITLE-ko ์งˆ์˜์˜ 27.8% ๊ฐ€ ์˜ˆ์ „ 64 ํ† ํฐ ํ•œ๋„๋ฅผ ๋„˜์—ˆ๊ณ , 77 ํ† ํฐ์„ ๋„˜๋Š” ์งˆ์˜๋Š” 10.8% ์ž…๋‹ˆ๋‹ค. K2 ๋Š” ์ด๋Ÿฐ ์งˆ์˜์—์„œ ์•ž 57 ํ† ํฐ๊ณผ ๋ 20 ํ† ํฐ์„ ๋‚จ๊ธฐ๋ฏ€๋กœ ๋์— ๋ถ™๋Š” ์ƒ‰์ƒยทํฌ๊ธฐยท์ˆ˜๋Ÿ‰ ํ‘œํ˜„์ด ์ž˜๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
73
 
74
  ## ์ฃผ์žฅํ•˜์ง€ ์•Š๋Š” ๊ฒƒ
 
43
 
44
  **๋ชจ๋ธ ๋‹จ๋… R@10**
45
 
46
+ ํ˜ธ์ถœ ๋ฐฉ์‹ ์„ธ ๊ฐ€์ง€๋ฅผ ๋‚˜๋ˆ  ์ ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์งˆ์˜๋ฅผ E0 ๋กœ ๋ณด๋‚ธ ๊ฐ’, ๋ชจ๋“  ์งˆ์˜๋ฅผ K2 ๋กœ ๋ณด๋‚ธ ๊ฐ’, ๊ธฐ๋ณธ ์ž๋™ ๋ผ์šฐํŒ…(ํ•œ๊ธ€์ด ์žˆ์œผ๋ฉด K2, ์—†์œผ๋ฉด E0) ๊ฐ’์ž…๋‹ˆ๋‹ค.
47
+
48
+ | ์งˆ์˜ ์„ธํŠธ | n | E0 ๋งŒ (`route="en"`) | ์ด์ „ K1 | **K2 ๋งŒ (`route="ko"`)** | ์ž๋™ ๋ผ์šฐํŒ…(๊ธฐ๋ณธ) | ๊ต์‚ฌ(1.5 GB) |
49
  |---|---|---|---|---|---|---|
50
  | Q-TITLE-ko: ํ•œ๊ตญ์–ด ์ƒํ’ˆ๋ช…(ABO ํ•œ๊ตญ์–ด ๋ฆฌ์ŠคํŒ…) | 889 | 0.368 | 0.231 | **0.645** | 0.643 | 0.682 |
51
  | kobench: ABO ์†์„ฑ ๊ธฐ๋ฐ˜ ํ•œ๊ตญ์–ด ํ…œํ”Œ๋ฆฟ ์งˆ์˜ | 173 | 0.104 | 0.069 | **0.318** | 0.318 | 0.491 |
 
56
  - **ํ•œ๊ตญ์–ด ์ƒํ’ˆ๋ช…์—์„œ K2 ๋Š” E0 ๋ณด๋‹ค R@10 ์ด 0.277 ๋†’์Šต๋‹ˆ๋‹ค** [+0.243, +0.309](์ƒํ’ˆ ์œ ํ˜• ๊ตฐ์ง‘ 150๊ฐœ). ์ด๋ฒˆ ๊ณต๊ฐœ์˜ ๊ทผ๊ฑฐ๊ฐ€ ์ด ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.
57
  - kobench ์—์„œ๋„ K2 โˆ’ E0 ๋Š” +0.214 [+0.113, +0.315] ์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ ์ƒํ’ˆ ์œ ํ˜• ๊ตฐ์ง‘์ด 24๊ฐœ๋ฟ์ด๋ผ ํŒ์ • ๊ธฐ์ค€(30๊ฐœ)์— ๋ชป ๋ฏธ์ณ "๋ถˆํ™•์‹ค"๋กœ ๋‘ก๋‹ˆ๋‹ค.
58
  - K-Fashion(ํ•œ๊ตญ์–ด ํŒจ์…˜ ์†์„ฑ, ํ•™์Šต ๋ถ„ํฌ ์•ˆ, P@1, n = 626)์—์„œ๋Š” E0 0.005, ์ด์ „ K1 **0.403**, K2 0.289 ์ž…๋‹ˆ๋‹ค. **์ด ์„ธํŠธ์—์„œ๋Š” K2 ๊ฐ€ ์ด์ „ ํ•œ๊ตญ์–ด ๋ชจ๋ธ๋ณด๋‹ค ๋‚ฎ์Šต๋‹ˆ๋‹ค.**
59
+ - Q-TITLE-ko ์—์„œ ์ž๋™ ๋ผ์šฐํŒ…(0.643)์ด K2 ๋งŒ ์“ด ๊ฐ’(0.645)๋ณด๋‹ค ์กฐ๊ธˆ ๋‚ฎ์Šต๋‹ˆ๋‹ค. ABO ํ•œ๊ตญ์–ด ๋ฆฌ์ŠคํŒ… ์ด๋ฆ„ ๊ฐ€์šด๋ฐ ์‹ค์ œ๋กœ๋Š” ์˜์–ด๋กœ ์ ํžŒ ๊ฒƒ(ํ•œ๊ธ€ ์—†์Œ)์ด E0 ๋กœ ๊ฐ€๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
60
  - ์˜์–ด ์ƒํ’ˆ๋ช…์€ ๊ทธ๋Œ€๋กœ์ด๊ณ (โˆ’0.001 [โˆ’0.009, +0.008]), ๋ฆฌ์ŠคํŒ… ํ‚ค์›Œ๋“œ๋Š” ์กฐ๊ธˆ ๋‚ฎ์•„์กŒ์Šต๋‹ˆ๋‹ค(โˆ’0.018 [โˆ’0.030, โˆ’0.006]). ์‹ค์ œ ์‡ผํ•‘ ๊ฒ€์ƒ‰์–ด๋Š” ์˜คํžˆ๋ ค ์˜ฌ๋ž์Šต๋‹ˆ๋‹ค(+0.100 [+0.059, +0.140]). ๊ธฐ๋ณธ ๋ผ์šฐํŒ…์€ ์˜์–ด๋ฅผ E0 ๋กœ ๋ณด๋‚ด๋ฏ€๋กœ ์˜์–ด ์ˆ˜์น˜๋Š” E0 ์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์˜์–ด ํŠธ๋ž˜ํ”ฝ์ด ESCI ์™€ ๋น„์Šทํ•˜๋‹ค๋ฉด `route="ko"` ๋ฅผ ๊ฒ€ํ† ํ•ด ๋ณด์„ธ์š”.
61
 
62
  **BM25 ์™€ ๊ฒฐํ•ฉ**(๋‘ ์ƒ์œ„ 10๊ฐœ ๋ชฉ๋ก์„ RRF, k = 60 ์œผ๋กœ ํ•ฉ์นจ. ํ•œ๊ตญ์–ด ์งˆ์˜์˜ BM25 ์ƒ‰์ธ์€ ์ƒํ’ˆ์˜ ํ•œ๊ตญ์–ดยท์˜์–ด ์ด๋ฆ„)
63
 
64
+ | ์งˆ์˜ ์„ธํŠธ | BM25 R@1 | BM25 R@10 | BM25 + K2 ๋งŒ R@1 | BM25 + K2 ๋งŒ R@10 |
65
  |---|---|---|---|---|
66
  | Q-TITLE-ko | 0.911 | 0.993 | 0.703 | 0.991 |
67
  | kobench | 0.139 | 0.405 | 0.179 | 0.468 |
 
72
  - ESCI(์˜์–ด)์—์„œ๋Š” K2 ๋ฅผ ๊ฒฐํ•ฉํ•˜๋ฉด BM25 ๋ณด๋‹ค R@10 ์ด +0.141 [+0.105, +0.181] ์˜ค๋ฆ…๋‹ˆ๋‹ค.
73
  - ์ด ๋ชจ๋ธ์˜ ์“ฐ์ž„์ƒˆ๋Š” ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค. ์ƒํ’ˆ ๋ชฉ๋ก์— ๋งž๋Š” ํ•œ๊ตญ์–ด ํ…์ŠคํŠธ๊ฐ€ ์—†์„ ๋•Œ ํ•œ๊ตญ์–ด ๊ฒ€์ƒ‰์–ด๋กœ **์ƒํ’ˆ ์ด๋ฏธ์ง€**๋ฅผ ์ฐพ๋Š” ๊ฒƒ, ๊ทธ๋ฆฌ๊ณ  ํ‚ค์›Œ๋“œ ๊ฒ€์ƒ‰์˜ ์ƒ์œ„ 10๊ฐœ ํ›„๋ณด๋ฅผ ๋„“ํžˆ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฒซ ๊ฒฐ๊ณผ๋ฅผ ์ •ํ•˜๋Š” ์ˆœ์œ„ ๋ชจ๋ธ๋กœ ์“ฐ์ง€ ๋งˆ์„ธ์š”.
74
 
75
+ ์„ค์น˜ ํ™•์ธ(์„ฑ๋Šฅ ์ง€ํ‘œ ์•„๋‹˜): ๊นจ๋—ํ•œ ํ™˜๊ฒฝ์—์„œ ์ด ์ €์žฅ์†Œ๋ฅผ ๋‚ด๋ ค๋ฐ›์•„ ABO ๊ฐœ๋ฐœ ์ƒํ’ˆ 200๊ฐœ(์ƒํ’ˆ ๊ตฐ์ง‘๋งˆ๋‹ค 1๊ฐœ)๋กœ EmbeddingGemma 2 ์ƒ‰์ธ์„ ๋งŒ๋“ค๊ณ , ๊ทธ ์ƒํ’ˆ๋“ค์˜ ํ•œ๊ตญ์–ด ์ด๋ฆ„์œผ๋กœ ๊ธฐ๋ณธ ๋ผ์šฐํŒ… ๊ฒ€์ƒ‰์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. 200๊ฐœ ์ค‘ 194๊ฐœ๋Š” ์ด๋ฆ„์— ํ•œ๊ธ€์ด ์žˆ์–ด K2 ๋กœ ๊ฐ”๊ณ , ๋‚˜๋จธ์ง€ 6๊ฐœ๋Š” ํ•œ๊ตญ์–ด ๋ฆฌ์ŠคํŒ… ์ด๋ฆ„์ด ์˜์–ด๋กœ ์ ํ˜€ ์žˆ์–ด(์˜ˆ: "AmazonBasics Hanging Planter, Square") E0 ๋กœ ๊ฐ”์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” R@1 0.63, R@10 0.95 ์ž…๋‹ˆ๋‹ค. ์˜์–ด ์ด๋ฆ„(์ „๋ถ€ E0)์œผ๋กœ๋Š” R@1 0.83, R@10 1.00 ์ž…๋‹ˆ๋‹ค. ํ’€์ด ์ž‘์•„์„œ ์ด ๊ฐ’์€ ํŒŒ์ผ์ด ์ œ๋Œ€๋กœ ๋ถˆ๋Ÿฌ์™€์ง€๊ณ  ๊ฒ€์ƒ‰๋œ๋‹ค๋Š” ํ™•์ธ์ผ ๋ฟ์ž…๋‹ˆ๋‹ค. ํ’ˆ์งˆ์€ ์œ„ ํ‘œ๋ฅผ ๋ณด์„ธ์š”.
76
+
77
  ์งˆ์˜ ๊ธธ์ด: Q-TITLE-ko ์งˆ์˜์˜ 27.8% ๊ฐ€ ์˜ˆ์ „ 64 ํ† ํฐ ํ•œ๋„๋ฅผ ๋„˜์—ˆ๊ณ , 77 ํ† ํฐ์„ ๋„˜๋Š” ์งˆ์˜๋Š” 10.8% ์ž…๋‹ˆ๋‹ค. K2 ๋Š” ์ด๋Ÿฐ ์งˆ์˜์—์„œ ์•ž 57 ํ† ํฐ๊ณผ ๋ 20 ํ† ํฐ์„ ๋‚จ๊ธฐ๋ฏ€๋กœ ๋์— ๋ถ™๋Š” ์ƒ‰์ƒยทํฌ๊ธฐยท์ˆ˜๋Ÿ‰ ํ‘œํ˜„์ด ์ž˜๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
78
 
79
  ## ์ฃผ์žฅํ•˜์ง€ ์•Š๋Š” ๊ฒƒ
README.md CHANGED
@@ -79,7 +79,9 @@ All numbers are from development sets. Two of them, Q-TITLE-ko and kobench, were
79
 
80
  **Model alone, R@10**
81
 
82
- | Query set | n | E0 | earlier K1 | **K2** | Routed (default) | Teacher (1.5 GB) |
 
 
83
  |---|---|---|---|---|---|---|
84
  | Q-TITLE-ko: Korean product names (ABO Korean listings) | 889 | 0.368 | 0.231 | **0.645** | 0.643 | 0.682 |
85
  | kobench: Korean template queries over ABO attributes | 173 | 0.104 | 0.069 | **0.318** | 0.318 | 0.491 |
@@ -90,11 +92,12 @@ All numbers are from development sets. Two of them, Q-TITLE-ko and kobench, were
90
  - **Korean product names: K2 โˆ’ E0 = +0.277 [+0.243, +0.309] R@10** (150 product-type clusters). This is the result the release rests on.
91
  - kobench: K2 โˆ’ E0 = +0.214 [+0.113, +0.315], but kobench has only 24 product-type clusters, below the 30 we require for a firm call, so we label it uncertain.
92
  - K-Fashion (Korean fashion attributes, in-distribution, P@1, n = 626): E0 0.005 ยท earlier K1 **0.403** ยท K2 0.289. **K2 is worse than the earlier Korean model here.**
 
93
  - English: unchanged on product names (โˆ’0.001 [โˆ’0.009, +0.008]), lower on listing keywords (โˆ’0.018 [โˆ’0.030, โˆ’0.006]), higher on real shopping queries (+0.100 [+0.059, +0.140]). The default routing sends English queries to E0, so it gives E0's English numbers. Use `route="ko"` if your English traffic looks like ESCI.
94
 
95
  **Fused with BM25** (reciprocal rank fusion of the two top-10 lists, k = 60; BM25 over the product title text, Korean + English names for Korean queries)
96
 
97
- | Query set | BM25 R@1 | BM25 R@10 | BM25 + K2 R@1 | BM25 + K2 R@10 |
98
  |---|---|---|---|---|
99
  | Q-TITLE-ko | 0.911 | 0.993 | 0.703 | 0.991 |
100
  | kobench | 0.139 | 0.405 | 0.179 | 0.468 |
@@ -105,6 +108,8 @@ All numbers are from development sets. Two of them, Q-TITLE-ko and kobench, were
105
  - ESCI (English): fusion with K2 adds +0.141 [+0.105, +0.181] R@10 over BM25.
106
  - What the model is for: searching **product images** with Korean queries when the catalog lacks matching Korean text, and widening a keyword search's top 10. It is not a first-result ranker.
107
 
 
 
108
  Query length: on Q-TITLE-ko, 27.8% of queries exceeded the old 64-token limit and 10.8% exceed 77. K2 keeps the first 57 and last 20 tokens of those, so trailing colour, size or quantity words are not cut off.
109
 
110
  ## Not claimed
 
79
 
80
  **Model alone, R@10**
81
 
82
+ Three ways to call the release are reported separately: E0 for every query, K2 for every query, and the default routing (Hangul โ†’ K2, otherwise E0).
83
+
84
+ | Query set | n | E0 only (`route="en"`) | earlier K1 | **K2 only (`route="ko"`)** | Automatic routing (default) | Teacher (1.5 GB) |
85
  |---|---|---|---|---|---|---|
86
  | Q-TITLE-ko: Korean product names (ABO Korean listings) | 889 | 0.368 | 0.231 | **0.645** | 0.643 | 0.682 |
87
  | kobench: Korean template queries over ABO attributes | 173 | 0.104 | 0.069 | **0.318** | 0.318 | 0.491 |
 
92
  - **Korean product names: K2 โˆ’ E0 = +0.277 [+0.243, +0.309] R@10** (150 product-type clusters). This is the result the release rests on.
93
  - kobench: K2 โˆ’ E0 = +0.214 [+0.113, +0.315], but kobench has only 24 product-type clusters, below the 30 we require for a firm call, so we label it uncertain.
94
  - K-Fashion (Korean fashion attributes, in-distribution, P@1, n = 626): E0 0.005 ยท earlier K1 **0.403** ยท K2 0.289. **K2 is worse than the earlier Korean model here.**
95
+ - Automatic routing on Q-TITLE-ko (0.643) is slightly below K2 only (0.645) because some ABO Korean-listing names are actually English text (no Hangul) and are sent to E0.
96
  - English: unchanged on product names (โˆ’0.001 [โˆ’0.009, +0.008]), lower on listing keywords (โˆ’0.018 [โˆ’0.030, โˆ’0.006]), higher on real shopping queries (+0.100 [+0.059, +0.140]). The default routing sends English queries to E0, so it gives E0's English numbers. Use `route="ko"` if your English traffic looks like ESCI.
97
 
98
  **Fused with BM25** (reciprocal rank fusion of the two top-10 lists, k = 60; BM25 over the product title text, Korean + English names for Korean queries)
99
 
100
+ | Query set | BM25 R@1 | BM25 R@10 | BM25 + K2 only R@1 | BM25 + K2 only R@10 |
101
  |---|---|---|---|---|
102
  | Q-TITLE-ko | 0.911 | 0.993 | 0.703 | 0.991 |
103
  | kobench | 0.139 | 0.405 | 0.179 | 0.468 |
 
108
  - ESCI (English): fusion with K2 adds +0.141 [+0.105, +0.181] R@10 over BM25.
109
  - What the model is for: searching **product images** with Korean queries when the catalog lacks matching Korean text, and widening a keyword search's top 10. It is not a first-result ranker.
110
 
111
+ Install check (not a benchmark): in a clean environment we downloaded this repository, built an EmbeddingGemma 2 index of 200 ABO development products (one per product cluster) and searched it with their Korean names using the default routing. 194 of the 200 names contain Hangul and went to K2; the other 6 Korean-listing names are English text (for example "AmazonBasics Hanging Planter, Square") and went to E0. Result: R@1 0.63, R@10 0.95. With English names (all routed to E0): R@1 0.83, R@10 1.00. The pool is tiny, so these numbers only show the files load and search correctly; use the table above for quality.
112
+
113
  Query length: on Q-TITLE-ko, 27.8% of queries exceeded the old 64-token limit and 10.8% exceed 77. K2 keeps the first 57 and last 20 tokens of those, so trailing colour, size or quantity words are not cut off.
114
 
115
  ## Not claimed