AsyiraFitri commited on Apr 28

Commit

e6d0ed4

1 Parent(s): 63b3a89

included the remaining embeddings that wasnt pushed previously

Files changed (17) hide show

tabacco3482_benchmarks/embeddings/test_emb_LayoutLMv3_Base.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:1c5643503131835d428f6e639d12b5e4d96c3c1b4a07014b62995931ea12582a
+size 2141312

tabacco3482_benchmarks/embeddings/test_emb_LayoutLMv3_Base_OCR.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:f13981c9ca68a17ef3a4f709ec46fb3eae63911fb78f2d4fe9f4462c7c1a70dd
+size 2141312

tabacco3482_benchmarks/embeddings/test_emb_LayoutLMv3_Large.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:e08ac12b4cdf2fee5ef53a4c9c3c08f7e211136c0ff9b038db70cc7dd4d3a8e0
+size 2855040

tabacco3482_benchmarks/embeddings/test_emb_LayoutLMv3_Large_OCR.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:dfbd6a16e147693a1cc960c4b8f283163fa877a825c42f89394407ad7a5302f7
+size 2855040

tabacco3482_benchmarks/embeddings/test_emb_SigLIP2_SO400M_finetuned.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:ac01150d7fd9b83a62451dc8d57bae8c950b2b49bf2fe0601647a7604ef6d917
+size 6423680

tabacco3482_benchmarks/embeddings/test_emb_SigLIP2_SO400M_hnm.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:f9318fc8ca4d997300ee3cadc77c1b5cbaeff6977bb868aec54903a983da31ef
+size 6423680

tabacco3482_benchmarks/embeddings/test_emb_SigLIP2_SO400M_patch14-384_clspatch.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:4a29335833760cfd179f0feea0504d3ea3fbcf03325be30a87bad642a5002a86
+size 6423680

tabacco3482_benchmarks/embeddings/test_labels.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:330e9f6a7ede728299b23e7c25b6411c83ce1a31cad3c07abaee45df8791201a
+size 2916

tabacco3482_benchmarks/embeddings/train_emb_LayoutLMv3_Base.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:dec8f70212e6fe2044e40d34fa5acecfb0e248914c9bcb6c17b941f470b65485
+size 8555648

tabacco3482_benchmarks/embeddings/train_emb_LayoutLMv3_Base_OCR.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:e315e0a24a2cbe64e38c79cbce62dd010d4250d38e89545f15cae0b7ed3f8b41
+size 8555648

tabacco3482_benchmarks/embeddings/train_emb_LayoutLMv3_Large.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:01ba4ab1b3f16a6ff32536a0d7cf727ea0e7430477627d4752210485d228cb51
+size 11407488

tabacco3482_benchmarks/embeddings/train_emb_LayoutLMv3_Large_OCR.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:5ae911919562746b5340c0d99abd3b9860ef5f69142a3bd7899ef62b95451e86
+size 11407488

tabacco3482_benchmarks/embeddings/train_emb_SigLIP2_SO400M_finetuned.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:0c534547cfda616fbca528531c90c67c561043b0c67589dae0ed84af301fa140
+size 25666688

tabacco3482_benchmarks/embeddings/train_emb_SigLIP2_SO400M_hnm.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:ab7901920441ed77d4605d6204d12097d232b1bcf1a03656c7f34de680165de7
+size 25666688

tabacco3482_benchmarks/embeddings/train_emb_SigLIP2_SO400M_patch14-384_clspatch.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:eb8c94b515587aa00d6789211cb2771aba2c7639f287917ada3cd8451587d3b7
+size 25666688

tabacco3482_benchmarks/embeddings/train_labels.npy ADDED Viewed

+version https://git-lfs.github.com/spec/v1
+oid sha256:53097fe8c9478e78db195192110785ff77283f35c33fb905ef664677fd41648d
+size 11268

tabacco3482_benchmarks/figures/layoutlmv3/result_analysis_layoutlmv3.md CHANGED Viewed

@@ -180,7 +180,7 @@ Both Large OCR and Large OCR + center achieve 0% error on Resume. The two-column
 Comparing Base (40.03% kNN@1) with Large (61.84%) shows a 21.8 pp gain from scaling. But comparing Large (61.84%) with Large OCR (79.77%) shows a 17.9 pp gain just from enabling OCR text — comparable to the full Base→Large scale-up. For a text-rich dataset like Tobacco-3482, OCR integration is at least as valuable as model scale.
 ### Bad Cases Visualisation
-→ [Bad Cases Top10 Visualisation](badcases/)
 ---
 ## 3. Recommendations

 Comparing Base (40.03% kNN@1) with Large (61.84%) shows a 21.8 pp gain from scaling. But comparing Large (61.84%) with Large OCR (79.77%) shows a 17.9 pp gain just from enabling OCR text — comparable to the full Base→Large scale-up. For a text-rich dataset like Tobacco-3482, OCR integration is at least as valuable as model scale.
 ### Bad Cases Visualisation
+→ [ Large Bad Cases Top10 Visualisation](badcases/)
 ---
 ## 3. Recommendations