Hugging Face's logo

Datasets:
AI-EcoNet
/
HUGO-Bench-Paper-Reproducibility

Tasks:
Image Classification
Zero-Shot Classification
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
ArXiv:
Tags:
biology
ecology
wildlife
camera-traps
vision-transformers
clustering
DOI:
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
HugoMarkoff commited on
Commit
fe13d81
·
verified ·
1 Parent(s): 9883b1a

Add BioCLIP 2.5 embeddings and benchmark artifacts

Browse files
Files changed (1) hide show
  1. bioclip2_5_benchmark/CURRENT_LEADERBOARDS.md +118 -0
bioclip2_5_benchmark/CURRENT_LEADERBOARDS.md ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Current Leaderboards
2
+
3
+ Source:
4
+ - Official HF primary benchmark: `AI-EcoNet/HUGO-Bench-Paper-Reproducibility` config `primary_benchmarking`
5
+ - Local BioCLIP 2.5 run: `results/benchmark_bioclip2_5_full_20260326_185250`
6
+
7
+ Notes:
8
+ - Official leaderboard below uses split-balanced mean V-measure = mean(Aves mean, Mammalia mean).
9
+ - `n_rows` can differ by method due to repeated stochastic runs (e.g., TSNE/UMAP).
10
+ - Local BioCLIP 2.5 score uses overall V-measure from the full-data run (Aves+Mammals together).
11
+ - Combined ranking therefore mixes two score types (documented per row in CSV).
12
+
13
+ ## How Scores Are Computed
14
+
15
+ - Official config score: split_balanced_v_mean = (mean_v_aves + mean_v_mammalia) / 2
16
+ - Local BioCLIP 2.5 score: overall_v_measure from one full-data run
17
+ - Top models (mean): arithmetic mean of the config-level score over all configs listed for that model
18
+
19
+ ## Official HF: Best Config Per Model
20
+
21
+ | Model | Reduction | Clustering | Split-balanced V | Aves V | Mammalia V | n_rows |
22
+ |---|---|---|---:|---:|---:|---:|
23
+ | dinov3_vith16plus | TSNE | HIERARCHICAL_K30 | 0.9576 | 0.9612 | 0.9541 | 200 |
24
+ | dinov2_vitg14 | TSNE | HIERARCHICAL_K30 | 0.8790 | 0.9026 | 0.8553 | 200 |
25
+ | bioclip2_vitl14 | TSNE | HIERARCHICAL_K45 | 0.7508 | 0.7814 | 0.7202 | 200 |
26
+ | clip_vitl14 | TSNE | HIERARCHICAL_K45 | 0.7116 | 0.7287 | 0.6944 | 200 |
27
+ | siglip_vitb16 | TSNE | HIERARCHICAL_K45 | 0.6864 | 0.6988 | 0.6740 | 200 |
28
+
29
+ ## Official HF: Global Top 20 Configurations
30
+
31
+ | Rank | Model | Reduction | Clustering | Split-balanced V | Aves V | Mammalia V | n_rows |
32
+ |---:|---|---|---|---:|---:|---:|---:|
33
+ | 1 | dinov3_vith16plus | TSNE | HIERARCHICAL_K30 | 0.9576 | 0.9612 | 0.9541 | 200 |
34
+ | 2 | dinov3_vith16plus | TSNE | GMM_K30 | 0.9559 | 0.9580 | 0.9538 | 200 |
35
+ | 3 | dinov3_vith16plus | UMAP | HIERARCHICAL_K30 | 0.9545 | 0.9596 | 0.9493 | 200 |
36
+ | 4 | dinov3_vith16plus | UMAP | GMM_K30 | 0.9528 | 0.9562 | 0.9494 | 200 |
37
+ | 5 | dinov3_vith16plus | TSNE | HDBSCAN | 0.9435 | 0.9476 | 0.9393 | 200 |
38
+ | 6 | dinov3_vith16plus | TSNE | HIERARCHICAL_K45 | 0.9227 | 0.9233 | 0.9222 | 200 |
39
+ | 7 | dinov3_vith16plus | UMAP | HIERARCHICAL_K45 | 0.9223 | 0.9233 | 0.9214 | 200 |
40
+ | 8 | dinov3_vith16plus | TSNE | GMM_K45 | 0.9219 | 0.9234 | 0.9203 | 200 |
41
+ | 9 | dinov3_vith16plus | UMAP | GMM_K45 | 0.9212 | 0.9224 | 0.9199 | 200 |
42
+ | 10 | dinov3_vith16plus | UMAP | HDBSCAN | 0.9031 | 0.9167 | 0.8896 | 200 |
43
+ | 11 | dinov2_vitg14 | TSNE | HIERARCHICAL_K30 | 0.8790 | 0.9026 | 0.8553 | 200 |
44
+ | 12 | dinov2_vitg14 | TSNE | GMM_K30 | 0.8767 | 0.9003 | 0.8530 | 200 |
45
+ | 13 | dinov2_vitg14 | UMAP | HIERARCHICAL_K30 | 0.8743 | 0.9003 | 0.8483 | 200 |
46
+ | 14 | dinov2_vitg14 | TSNE | HDBSCAN | 0.8731 | 0.9028 | 0.8433 | 200 |
47
+ | 15 | dinov2_vitg14 | UMAP | GMM_K30 | 0.8729 | 0.9006 | 0.8453 | 200 |
48
+ | 16 | dinov2_vitg14 | TSNE | HIERARCHICAL_K45 | 0.8661 | 0.8833 | 0.8489 | 200 |
49
+ | 17 | dinov2_vitg14 | TSNE | GMM_K45 | 0.8628 | 0.8814 | 0.8442 | 200 |
50
+ | 18 | dinov2_vitg14 | UMAP | HIERARCHICAL_K45 | 0.8584 | 0.8790 | 0.8379 | 200 |
51
+ | 19 | dinov2_vitg14 | UMAP | GMM_K45 | 0.8556 | 0.8777 | 0.8335 | 200 |
52
+ | 20 | dinov3_vith16plus | TSNE | GMM_K90 | 0.8458 | 0.8462 | 0.8453 | 200 |
53
+
54
+ ## Local BioCLIP 2.5 (Current Run, Non-Official Protocol)
55
+
56
+ | Rank | Reduction | Clustering | Config | V-Measure | AMI | ARI | Outlier % | Samples |
57
+ |---:|---|---|---|---:|---:|---:|---:|---:|
58
+ | 1 | tsne | hdbscan | hdbscan_mcs150_ms50 | 0.8822 | 0.8814 | 0.7963 | 13.59 | 139067 |
59
+ | 2 | umap | hdbscan | hdbscan_mcs150_ms50 | 0.8590 | 0.8580 | 0.7149 | 15.28 | 139067 |
60
+ | 3 | umap | dbscan | dbscan_relaxed_ms20 | 0.7992 | 0.7971 | 0.4547 | 6.28 | 139067 |
61
+ | 4 | tsne | gmm | gmm_k60 | 0.7735 | 0.7728 | 0.5789 | 0.00 | 139067 |
62
+ | 5 | umap | gmm | gmm_k60 | 0.7727 | 0.7720 | 0.5476 | 0.00 | 139067 |
63
+ | 6 | tsne | dbscan | dbscan_relaxed_ms20 | 0.7717 | 0.7693 | 0.2577 | 2.90 | 139067 |
64
+ | 7 | pca | gmm | gmm_k60 | 0.3244 | 0.3223 | 0.0944 | 0.00 | 139067 |
65
+ | 8 | pca | hdbscan | hdbscan_mcs150_ms50 | 0.1804 | 0.1802 | 0.0130 | 1.06 | 139067 |
66
+ | 9 | pca | dbscan | dbscan_relaxed_ms20 | 0.0091 | 0.0056 | 0.0000 | 4.97 | 139067 |
67
+
68
+ ## Combined Snapshot: Top 20 Configurations (Official + BioCLIP 2.5)
69
+
70
+ | Rank | Source | Model | Reduction | Clustering | Config | Score | Score Type |
71
+ |---:|---|---|---|---|---|---:|---|
72
+ | 1 | official_hf | dinov3_vith16plus | TSNE | HIERARCHICAL_K30 | HIERARCHICAL_K30 | 0.9576 | split_balanced_v_mean |
73
+ | 2 | official_hf | dinov3_vith16plus | TSNE | GMM_K30 | GMM_K30 | 0.9559 | split_balanced_v_mean |
74
+ | 3 | official_hf | dinov3_vith16plus | UMAP | HIERARCHICAL_K30 | HIERARCHICAL_K30 | 0.9545 | split_balanced_v_mean |
75
+ | 4 | official_hf | dinov3_vith16plus | UMAP | GMM_K30 | GMM_K30 | 0.9528 | split_balanced_v_mean |
76
+ | 5 | official_hf | dinov3_vith16plus | TSNE | HDBSCAN | HDBSCAN | 0.9435 | split_balanced_v_mean |
77
+ | 6 | official_hf | dinov3_vith16plus | TSNE | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.9227 | split_balanced_v_mean |
78
+ | 7 | official_hf | dinov3_vith16plus | UMAP | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.9223 | split_balanced_v_mean |
79
+ | 8 | official_hf | dinov3_vith16plus | TSNE | GMM_K45 | GMM_K45 | 0.9219 | split_balanced_v_mean |
80
+ | 9 | official_hf | dinov3_vith16plus | UMAP | GMM_K45 | GMM_K45 | 0.9212 | split_balanced_v_mean |
81
+ | 10 | official_hf | dinov3_vith16plus | UMAP | HDBSCAN | HDBSCAN | 0.9031 | split_balanced_v_mean |
82
+ | 11 | local_bioclip2_5_full | bioclip2_5_vith14 | TSNE | HDBSCAN | hdbscan_mcs150_ms50 | 0.8822 | overall_v_measure |
83
+ | 12 | official_hf | dinov2_vitg14 | TSNE | HIERARCHICAL_K30 | HIERARCHICAL_K30 | 0.8790 | split_balanced_v_mean |
84
+ | 13 | official_hf | dinov2_vitg14 | TSNE | GMM_K30 | GMM_K30 | 0.8767 | split_balanced_v_mean |
85
+ | 14 | official_hf | dinov2_vitg14 | UMAP | HIERARCHICAL_K30 | HIERARCHICAL_K30 | 0.8743 | split_balanced_v_mean |
86
+ | 15 | official_hf | dinov2_vitg14 | TSNE | HDBSCAN | HDBSCAN | 0.8731 | split_balanced_v_mean |
87
+ | 16 | official_hf | dinov2_vitg14 | UMAP | GMM_K30 | GMM_K30 | 0.8729 | split_balanced_v_mean |
88
+ | 17 | official_hf | dinov2_vitg14 | TSNE | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.8661 | split_balanced_v_mean |
89
+ | 18 | official_hf | dinov2_vitg14 | TSNE | GMM_K45 | GMM_K45 | 0.8628 | split_balanced_v_mean |
90
+ | 19 | local_bioclip2_5_full | bioclip2_5_vith14 | UMAP | HDBSCAN | hdbscan_mcs150_ms50 | 0.8590 | overall_v_measure |
91
+ | 20 | official_hf | dinov2_vitg14 | UMAP | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.8584 | split_balanced_v_mean |
92
+
93
+ ## Combined Snapshot: Best Config Per Model
94
+
95
+ | Model | Source | Reduction | Clustering | Config | Score | Score Type |
96
+ |---|---|---|---|---|---:|---|
97
+ | dinov3_vith16plus | official_hf | TSNE | HIERARCHICAL_K30 | HIERARCHICAL_K30 | 0.9576 | split_balanced_v_mean |
98
+ | bioclip2_5_vith14 | local_bioclip2_5_full | TSNE | HDBSCAN | hdbscan_mcs150_ms50 | 0.8822 | overall_v_measure |
99
+ | dinov2_vitg14 | official_hf | TSNE | HIERARCHICAL_K30 | HIERARCHICAL_K30 | 0.8790 | split_balanced_v_mean |
100
+ | bioclip2_vitl14 | official_hf | TSNE | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.7508 | split_balanced_v_mean |
101
+ | clip_vitl14 | official_hf | TSNE | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.7116 | split_balanced_v_mean |
102
+ | siglip_vitb16 | official_hf | TSNE | HIERARCHICAL_K45 | HIERARCHICAL_K45 | 0.6864 | split_balanced_v_mean |
103
+
104
+ ## Top Models (Mean Score)
105
+
106
+ | Rank | Model | Mean Score | #Configs | Sources |
107
+ |---:|---|---:|---:|---|
108
+ | 1 | dinov3_vith16plus | 0.6759 | 60 | official_hf |
109
+ | 2 | dinov2_vitg14 | 0.6479 | 60 | official_hf |
110
+ | 3 | bioclip2_5_vith14 | 0.5969 | 9 | local_bioclip2_5_full |
111
+ | 4 | bioclip2_vitl14 | 0.4849 | 60 | official_hf |
112
+ | 5 | clip_vitl14 | 0.4312 | 60 | official_hf |
113
+ | 6 | siglip_vitb16 | 0.4309 | 60 | official_hf |
114
+
115
+ ## Generated Files
116
+
117
+ - Official CSV: C:\Users\Hugo\Documents\zeroshot_clustering_animal_images\results\current_leaderboard_hf_primary.csv
118
+ - Combined CSV: C:\Users\Hugo\Documents\zeroshot_clustering_animal_images\results\current_leaderboard_combined_with_bioclip2_5.csv