Datasets:

cihatyldz
/

pusula-tr-robustness

model stringclasses 14 values	veri_seti stringclasses 4 values	siralamada bool 2 classes	bozulma stringclasses 6 values	temiz_ndcg@10 float64 0.13 0.97	bozuk_ndcg@10 float64 0.02 0.97	koruma_degisen float64 10.3 103	degisen_oran float64 0.16 1	fark_ci_alt float64 -0.76 0	fark_ci_ust float64 -0.72 0.04
qwen3-emb-4b	msmarco-tr	false	ascii_full	0.8813	0.8113	90.5	0.842	-0.0988	-0.0672
qwen3-emb-4b	msmarco-tr	false	ascii_partial	0.8813	0.8522	95	0.66	-0.0568	-0.0305
qwen3-emb-4b	msmarco-tr	false	upper	0.8813	0.8638	98	1	-0.0274	-0.0076
qwen3-emb-4b	msmarco-tr	false	typo	0.8813	0.8367	91.6	0.595	-0.0924	-0.0562
qwen3-emb-4b	msmarco-tr	false	colloquial	0.8813	0.8793	98.8	0.156	-0.0261	0.0061
qwen3-emb-4b	msmarco-tr	false	mobile	0.8813	0.7787	88	0.966	-0.1229	-0.0898
qwen3-emb-4b	quora-tr	true	ascii_full	0.9502	0.9234	97	0.949	-0.037	-0.0204
qwen3-emb-4b	quora-tr	true	ascii_partial	0.9502	0.9392	98.6	0.819	-0.0198	-0.0075
qwen3-emb-4b	quora-tr	true	upper	0.9502	0.9372	98.6	1	-0.0199	-0.0069
qwen3-emb-4b	quora-tr	true	typo	0.9502	0.9353	97.8	0.712	-0.0302	-0.0124
qwen3-emb-4b	quora-tr	true	colloquial	0.9502	0.9466	98.8	0.316	-0.0205	-0.0035
qwen3-emb-4b	quora-tr	true	mobile	0.9502	0.8894	93.6	0.998	-0.0746	-0.0483
qwen3-emb-4b	fiqa-tr	true	ascii_full	0.4271	0.3794	88.7	0.972	-0.0618	-0.0378
qwen3-emb-4b	fiqa-tr	true	ascii_partial	0.4271	0.411	95.9	0.907	-0.0272	-0.009
qwen3-emb-4b	fiqa-tr	true	upper	0.4271	0.398	93.2	1	-0.0399	-0.0181
qwen3-emb-4b	fiqa-tr	true	typo	0.4271	0.3971	91.3	0.785	-0.0518	-0.0254
qwen3-emb-4b	fiqa-tr	true	colloquial	0.4271	0.4252	99.1	0.483	-0.013	0.0048
qwen3-emb-4b	fiqa-tr	true	mobile	0.4271	0.3454	80.9	1	-0.0985	-0.0657
qwen3-emb-4b	dmm	true	ascii_full	0.8838	0.8676	98.1	0.978	-0.0234	-0.0101
qwen3-emb-4b	dmm	true	ascii_partial	0.8838	0.8794	99.5	0.929	-0.0092	-0.0002
qwen3-emb-4b	dmm	true	upper	0.8838	0.8809	99.7	0.999	-0.0081	0.0024
qwen3-emb-4b	dmm	true	typo	0.8838	0.8777	99.1	0.787	-0.0143	-0.0019
qwen3-emb-4b	dmm	true	colloquial	0.8838	0.8834	99.9	0.262	-0.0071	0.0039
qwen3-emb-4b	dmm	true	mobile	0.8838	0.8545	96.7	0.995	-0.0392	-0.0203
bge-m3	msmarco-tr	false	ascii_full	0.8872	0.7682	84.1	0.842	-0.1624	-0.1215
bge-m3	msmarco-tr	false	ascii_partial	0.8872	0.8273	89.8	0.66	-0.111	-0.0724
bge-m3	msmarco-tr	false	upper	0.8872	0.7855	88.5	1	-0.1188	-0.0849
bge-m3	msmarco-tr	false	typo	0.8872	0.8342	90.1	0.595	-0.1071	-0.0706
bge-m3	msmarco-tr	false	colloquial	0.8872	0.8837	97.4	0.156	-0.0414	-0.0055
bge-m3	msmarco-tr	false	mobile	0.8872	0.699	78.1	0.966	-0.2185	-0.1715
bge-m3	quora-tr	true	ascii_full	0.9591	0.9304	96.9	0.949	-0.0403	-0.0214
bge-m3	quora-tr	true	ascii_partial	0.9591	0.9459	98.3	0.819	-0.0243	-0.0089
bge-m3	quora-tr	true	upper	0.9591	0.9264	96.6	1	-0.0432	-0.023
bge-m3	quora-tr	true	typo	0.9591	0.9539	99.2	0.712	-0.0127	-0.0024
bge-m3	quora-tr	true	colloquial	0.9591	0.9591	100	0.316	-0.004	0.0032
bge-m3	quora-tr	true	mobile	0.9591	0.8962	93.4	0.998	-0.0769	-0.0495
bge-m3	fiqa-tr	true	ascii_full	0.3679	0.3097	84	0.972	-0.0729	-0.0468
bge-m3	fiqa-tr	true	ascii_partial	0.3679	0.3407	92.1	0.907	-0.042	-0.0176
bge-m3	fiqa-tr	true	upper	0.3679	0.2899	78.8	1	-0.0945	-0.063
bge-m3	fiqa-tr	true	typo	0.3679	0.34	90.5	0.785	-0.0494	-0.0228
bge-m3	fiqa-tr	true	colloquial	0.3679	0.365	98.5	0.483	-0.0122	-0.0001
bge-m3	fiqa-tr	true	mobile	0.3679	0.2734	74.3	1	-0.1109	-0.0774
bge-m3	dmm	true	ascii_full	0.8963	0.8657	96.5	0.978	-0.0414	-0.0217
bge-m3	dmm	true	ascii_partial	0.8963	0.8834	98.5	0.929	-0.0208	-0.0077
bge-m3	dmm	true	upper	0.8963	0.8746	97.6	0.999	-0.0299	-0.0138
bge-m3	dmm	true	typo	0.8963	0.8893	99	0.787	-0.0171	-0.0014
bge-m3	dmm	true	colloquial	0.8963	0.8952	99.5	0.262	-0.0096	-0.0002
bge-m3	dmm	true	mobile	0.8963	0.8605	96	0.995	-0.0469	-0.026
mursit-large-tr	msmarco-tr	false	ascii_full	0.9098	0.8574	93.1	0.842	-0.0762	-0.0497
mursit-large-tr	msmarco-tr	false	ascii_partial	0.9098	0.8748	94.1	0.66	-0.0678	-0.0398
mursit-large-tr	msmarco-tr	false	upper	0.9098	0.8003	88	1	-0.1268	-0.0938
mursit-large-tr	msmarco-tr	false	typo	0.9098	0.8538	89.6	0.595	-0.1145	-0.0745
mursit-large-tr	msmarco-tr	false	colloquial	0.9098	0.9087	99.2	0.156	-0.0194	0.0051
mursit-large-tr	msmarco-tr	false	mobile	0.9098	0.8007	87.6	0.966	-0.1306	-0.0961
mursit-large-tr	quora-tr	true	ascii_full	0.9445	0.9319	98.6	0.949	-0.0204	-0.0066
mursit-large-tr	quora-tr	true	ascii_partial	0.9445	0.9367	99	0.819	-0.0168	-0.0034
mursit-large-tr	quora-tr	true	upper	0.9445	0.9243	97.9	1	-0.027	-0.0134
mursit-large-tr	quora-tr	true	typo	0.9445	0.9307	98	0.712	-0.0288	-0.0104
mursit-large-tr	quora-tr	true	colloquial	0.9445	0.9425	99.4	0.316	-0.0134	-0.0002
mursit-large-tr	quora-tr	true	mobile	0.9445	0.8949	94.7	0.998	-0.062	-0.0383
mursit-large-tr	fiqa-tr	true	ascii_full	0.3095	0.2858	92.3	0.972	-0.0352	-0.0132
mursit-large-tr	fiqa-tr	true	ascii_partial	0.3095	0.2956	95.1	0.907	-0.0254	-0.0051
mursit-large-tr	fiqa-tr	true	upper	0.3095	0.2529	81.7	1	-0.0711	-0.0421
mursit-large-tr	fiqa-tr	true	typo	0.3095	0.2746	85.8	0.785	-0.0606	-0.0298
mursit-large-tr	fiqa-tr	true	colloquial	0.3095	0.3097	100.3	0.483	-0.008	0.0096
mursit-large-tr	fiqa-tr	true	mobile	0.3095	0.2463	79.6	1	-0.08	-0.0474
mursit-large-tr	dmm	true	ascii_full	0.8691	0.8586	98.8	0.978	-0.017	-0.0045
mursit-large-tr	dmm	true	ascii_partial	0.8691	0.865	99.5	0.929	-0.0084	-0.0001
mursit-large-tr	dmm	true	upper	0.8691	0.8588	98.8	0.999	-0.0167	-0.0039
mursit-large-tr	dmm	true	typo	0.8691	0.8583	98.5	0.787	-0.0213	-0.0069
mursit-large-tr	dmm	true	colloquial	0.8691	0.8691	100	0.262	-0.0017	0.0008
mursit-large-tr	dmm	true	mobile	0.8691	0.84	96.6	0.995	-0.0389	-0.0208
mE5-large	msmarco-tr	false	ascii_full	0.9045	0.7471	79.2	0.842	-0.2103	-0.165
mE5-large	msmarco-tr	false	ascii_partial	0.9045	0.8334	88.1	0.66	-0.1296	-0.0872
mE5-large	msmarco-tr	false	upper	0.9045	0.8108	89.6	1	-0.1095	-0.0788
mE5-large	msmarco-tr	false	typo	0.9045	0.8255	85.3	0.595	-0.1568	-0.1106
mE5-large	msmarco-tr	false	colloquial	0.9045	0.9006	97	0.156	-0.0518	-0.0014
mE5-large	msmarco-tr	false	mobile	0.9045	0.6717	73.3	0.966	-0.2653	-0.2186
mE5-large	quora-tr	true	ascii_full	0.9671	0.9337	96.4	0.949	-0.0452	-0.0259
mE5-large	quora-tr	true	ascii_partial	0.9671	0.9539	98.3	0.819	-0.0236	-0.0094
mE5-large	quora-tr	true	upper	0.9671	0.9497	98.2	1	-0.0241	-0.0111
mE5-large	quora-tr	true	typo	0.9671	0.9601	98.9	0.712	-0.0163	-0.0042
mE5-large	quora-tr	true	colloquial	0.9671	0.967	100	0.316	-0.0054	0.0034
mE5-large	quora-tr	true	mobile	0.9671	0.9078	93.9	0.998	-0.0724	-0.0473
mE5-large	fiqa-tr	true	ascii_full	0.3826	0.2827	73.6	0.972	-0.1197	-0.0859
mE5-large	fiqa-tr	true	ascii_partial	0.3826	0.3343	86.4	0.907	-0.0671	-0.0389
mE5-large	fiqa-tr	true	upper	0.3826	0.3104	81.1	1	-0.0878	-0.0566
mE5-large	fiqa-tr	true	typo	0.3826	0.3435	87.2	0.785	-0.0659	-0.0355
mE5-large	fiqa-tr	true	colloquial	0.3826	0.3748	95.9	0.483	-0.0284	-0.0055
mE5-large	fiqa-tr	true	mobile	0.3826	0.2262	59.1	1	-0.1765	-0.136
mE5-large	dmm	true	ascii_full	0.8932	0.8609	96.3	0.978	-0.0423	-0.024
mE5-large	dmm	true	ascii_partial	0.8932	0.8777	98.2	0.929	-0.0235	-0.0103
mE5-large	dmm	true	upper	0.8932	0.8809	98.6	0.999	-0.0185	-0.0068
mE5-large	dmm	true	typo	0.8932	0.8782	97.9	0.787	-0.0269	-0.0117
mE5-large	dmm	true	colloquial	0.8932	0.8928	99.8	0.262	-0.007	0.0032
mE5-large	dmm	true	mobile	0.8932	0.8463	94.7	0.995	-0.0588	-0.0357
trmteb-tr	msmarco-tr	false	ascii_full	0.8485	0.8321	97.7	0.842	-0.0283	-0.0119
trmteb-tr	msmarco-tr	false	ascii_partial	0.8485	0.8404	98.5	0.66	-0.0201	-0.0051
trmteb-tr	msmarco-tr	false	upper	0.8485	0.8321	98.1	1	-0.0236	-0.0099
trmteb-tr	msmarco-tr	false	typo	0.8485	0.7693	84.3	0.595	-0.1577	-0.1096

End of preview. Expand in Data Studio

🧭 Pusula — Türkçe retrieval dayanıklılık benchmark'ı

Türkçe arama/embedding modellerinin gerçek kullanıcı sorgularına dayanıklılığı: ASCII yazım, BÜYÜK HARF, yazım hatası, gündelik yazım ve hepsinin birleşimi (mobil). Yalnızca sorgu bozulur, dokümanlar temiz kalır.

🚀 Etkileşimli leaderboard: cihatyldz/pusula

Sonuçlar

Koruma = yalnızca bozulmanın değiştirdiği sorgularda bozuk/temiz ndcg@10 (%). Aynı grup numarası: fark istatistiksel olarak ayırt edilemez (veri seti dengeli eşleştirilmiş bootstrap, %95). Sıralama mobil sorgulardaki mutlak başarıya göre.

Ortalamalar ve sıralama yalnızca şu veri setlerinden hesaplanır: quora-tr, fiqa-tr, dmm. Sıralama dışı tutulan setler detail.csv içinde siralamada=False ile yer alır:

msmarco-tr: Test sorguları MS MARCO train setinin yeniden bölünmesiyle oluşturulmuş (train+dev+test = 532.761 = MS MARCO train ilgi etiketi sayısı); MS MARCO ile eğitilmiş modeller bu sorgu–pasaj çiftlerini eğitimde görmüş olabilir.

grup	model	aile	temiz_ndcg@10	mobil_ndcg@10	koruma_upper	koruma_ascii_full	ort_koruma
1	qwen3-emb-4b	Çok dilli	0.7537	0.6964	97.2	94.6	95.9
2	bge-m3	Çok dilli	0.7411	0.6767	91.0	92.5	93.9
3	mursit-large-tr	Türkçe	0.7077	0.6604	92.8	96.5	95.2
3	mE5-large	Çok dilli	0.7476	0.6601	92.7	88.8	91.9
3	trmteb-tr	Türkçe	0.6897	0.6523	98.0	98.0	96.4
4	cevher-nuve	Türkçe	0.6769	0.6387	84.8	95.1	92.9
5	mE5-base	Çok dilli	0.7102	0.5966	80.7	83.5	87.5
5	qwen3-emb-0.6b	Çok dilli	0.6839	0.5952	87.2	88.4	90.7
5	bm25-fold	Sözcüksel	0.6128	0.5858	100.0	100.0	97.1
5	mE5-small	Çok dilli	0.6872	0.5757	80.3	84.7	86.8
5	moganbert-embed	Türkçe	0.7221	0.5694	81.3	91.1	88.5
6	emrecan-bert-tr	Türkçe	0.6267	0.536	58.1	89.5	84.7
7	bm25	Sözcüksel	0.6108	0.4897	100.0	82.4	89.3
8	mMiniLM-L12	Çok dilli	0.6001	0.3591	16.6	68.6	67.1

İçerik

leaderboard.csv, detail.csv: model ve model × veri seti × bozulma düzeyinde sonuçlar
results/*.json: sorgu bazında skorlar (yalnızca sorgu ID'leri ve sayılar)
code/: benchmark'ı üreten ve değerlendiren modüller

Yeniden üretim

Benchmark metinleri lisans nedeniyle dağıtılmaz. code/ modülleri kaynak veri setlerinden aynı benchmark'ı üretir; parmak izi aaa2a0a453bb4f31 ile doğrulanır. Ayarlar: sorgu ≤1000. korpus ≤20.000. seed 42. bozulma seed 0.

Kaynaklar ve lisans

Skorlar ve kod: CC BY 4.0 / Apache-2.0
Kaynak veri: TR-MTEB (Baysan & Güngör, Findings of EMNLP 2025) msmarco-tr, quora-tr, fiqa-tr (orijinal veri setlerinin koşulları geçerlidir; MS MARCO yalnızca ticari olmayan araştırma) ve iletisim/dezenformasyon-bultenleri (İletişim Başkanlığı DMM bültenleri).

Sınırlamalar

Mutlak skorlar alt-örneklenmiş korpus nedeniyle TR-MTEB ile karşılaştırılamaz; iddia koruma oranlarıdır.
Üç veri seti makine çevirisi; bozulmalar kural tabanlı.

Katkılar: cevher-nuve sonucu @batuhanozkose tarafından önerildi. Katkı sonuçları Pusula kodu ve sabit model commit'iyle bağımsız olarak yeniden üretilir.

Geliştirici: Cihat Yıldız · @cihatyldz

Downloads last month: 128