Hugging Face's logo

Datasets:
Mergeability-2
/
beetle-merge-eval

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Dutch
German
Size:
1K - 10K
Tags:
model-merging
mergeability
blimp
multiblimp
evaluation
multilingual
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
suchirsalhan commited on
Commit
6f19ac1
·
verified ·
1 Parent(s): 1ce879f

Upload RESULTS_BEETLE_MERGE_EVAL.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. RESULTS_BEETLE_MERGE_EVAL.md +76 -30
RESULTS_BEETLE_MERGE_EVAL.md CHANGED
@@ -1,6 +1,6 @@
1
  # Beetle merged models vs their parents — evaluation report
2
 
3
- _Generated 2026-08-27 02:41 UTC. **PARTIAL RUN** — see the coverage table for exactly what is and is not scored._
4
 
5
 
6
  ## Headline findings
@@ -48,15 +48,40 @@ naive artefact to compare against. Rows show `beats naive 0/0`; read that as
48
 
49
  Full audit: `arm_identity_audit.csv`.
50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51
  ## Coverage
52
 
53
  | provenance_family | merges | rungs_published | rungs_scored | complete | partial | not_eval |
54
  |:--------------------|---------:|------------------:|---------------:|-----------:|----------:|-----------:|
55
- | arch | 1 | 6 | 1 | 0 | 1 | 0 |
56
  | humanscale-pair | 3 | 72 | 72 | 3 | 0 | 0 |
57
- | zoo-curriculum | 49 | 49 | 0 | 0 | 0 | 49 |
58
- | zoo-language | 111 | 692 | 38 | 2 | 3 | 106 |
59
- | zoo-mono_vs_joint | 50 | 50 | 1 | 1 | 0 | 49 |
60
 
61
  Per-merge coverage: `coverage_by_merge.csv`.
62
 
@@ -64,28 +89,31 @@ Per-merge coverage: `coverage_by_merge.csv`.
64
 
65
  | provenance_family | cells | merges | mean acc | mean Δ vs naive | beats naive | mean Δ vs better parent | beats better parent | mean Δ vs joint ceiling | beats ceiling |
66
  |:--------------------|--------:|---------:|-----------:|------------------:|:--------------|--------------------------:|:----------------------|--------------------------:|:----------------|
67
- | arch | 4 | 1 | 0.7052 | 0 | 0/4 | -0.033 | 1/4 | nan | 0/0 |
68
  | humanscale-pair | 240 | 3 | 0.5669 | 0.0055 | 42/240 | -0.1966 | 0/240 | -0.1416 | 4/240 |
69
- | zoo-language | 156 | 5 | 0.5425 | -0.0021 | 16/138 | -0.1943 | 8/156 | nan | 0/0 |
70
- | zoo-mono_vs_joint | 3 | 1 | 0.6651 | nan | 0/0 | -0.0477 | 0/3 | -0.046 | 0/3 |
 
71
 
72
  ### blimp_eng (chance 0.500)
73
 
74
  | provenance_family | cells | merges | mean acc | mean Δ vs naive | beats naive | mean Δ vs better parent | beats better parent | mean Δ vs joint ceiling | beats ceiling |
75
  |:--------------------|--------:|---------:|-----------:|------------------:|:--------------|--------------------------:|:----------------------|--------------------------:|:----------------|
76
- | arch | 1 | 1 | 0.557 | 0 | 0/1 | 0.0018 | 1/1 | nan | 0/0 |
77
  | humanscale-pair | 72 | 3 | 0.5242 | 0.0017 | 14/72 | -0.123 | 0/72 | -0.0345 | 1/72 |
78
- | zoo-language | 38 | 5 | 0.5444 | -0.0031 | 2/34 | -0.0407 | 8/38 | nan | 0/0 |
79
- | zoo-mono_vs_joint | 1 | 1 | 0.5284 | nan | 0/0 | -0.0173 | 0/1 | -0.0173 | 0/1 |
 
80
 
81
  ### multiblimp:eng (chance 0.500)
82
 
83
  | provenance_family | cells | merges | mean acc | mean Δ vs naive | beats naive | mean Δ vs better parent | beats better parent | mean Δ vs joint ceiling | beats ceiling |
84
  |:--------------------|--------:|---------:|-----------:|------------------:|:--------------|--------------------------:|:----------------------|--------------------------:|:----------------|
85
- | arch | 1 | 1 | 0.7169 | 0 | 0/1 | -0.013 | 0/1 | nan | 0/0 |
86
  | humanscale-pair | 72 | 3 | 0.5609 | 0.0093 | 13/72 | -0.2287 | 0/72 | -0.1512 | 0/72 |
87
- | zoo-language | 38 | 5 | 0.6443 | -0.0085 | 3/34 | -0.0704 | 0/38 | nan | 0/0 |
88
- | zoo-mono_vs_joint | 1 | 1 | 0.6468 | nan | 0/0 | -0.0299 | 0/1 | -0.0247 | 0/1 |
 
89
 
90
  ## Worked example — the three named `beetle-humanscale-*-eng` pairs
91
 
@@ -119,14 +147,14 @@ shown against both bars so the loss to the better parent stays visible):
119
 
120
  ### arch
121
 
122
- Scored **1 of 6** published rungs across 1 of 1 repos (4 model-benchmark cells).
123
 
124
  A single cross-architecture pair (pico x MoE).
125
 
126
- - **vs better parent**: 1/4 cells above; mean Δ -0.0330.
127
  - **vs joint ceiling**: no jointly-trained counterpart published for this family.
128
- - **vs naive**: 0/4 cells above; mean Δ +0.0000.
129
- - Mean accuracy 0.7052 against a chance floor of 0.500; 0% of cells fall within 2 points of chance.
130
 
131
  ### humanscale-pair
132
 
@@ -139,27 +167,38 @@ Three named pairs, all 24 rungs each, complete. The most informative family: it
139
  - **vs naive**: 42/240 cells above; mean Δ +0.0055.
140
  - Mean accuracy 0.5669 against a chance floor of 0.500; 20% of cells fall within 2 points of chance.
141
 
 
 
 
 
 
 
 
 
 
 
 
142
  ### zoo-language
143
 
144
- Scored **38 of 692** published rungs across 5 of 111 repos (156 model-benchmark cells).
145
 
146
  Largest family. Carries ties/average/task_arithmetic in naive, aligned and transport arms, so it is the only family with broad arm coverage — but the aligned arm is byte-identical to naive throughout.
147
 
148
- - **vs better parent**: 8/156 cells above; mean Δ -0.1943.
149
  - **vs joint ceiling**: no jointly-trained counterpart published for this family.
150
- - **vs naive**: 16/138 cells above; mean Δ -0.0021.
151
- - Mean accuracy 0.5425 against a chance floor of 0.500; 19% of cells fall within 2 points of chance.
152
 
153
  ### zoo-mono_vs_joint
154
 
155
- Scored **1 of 50** published rungs across 1 of 50 repos (3 model-benchmark cells).
156
 
157
  Parent B *is* the jointly-trained model, so the joint ceiling is built in and `d_ceiling` is meaningful here. Published rungs are `qmd_guided__aligned` and `qmd_guided__transport` only — no naive folder exists, so Δ vs naive is unavailable by construction.
158
 
159
- - **vs better parent**: 0/3 cells above; mean Δ -0.0477.
160
- - **vs joint ceiling**: 0/3 cells above; mean Δ -0.0460.
161
  - **vs naive**: **no naive artefact exists** for these repos, so this bar cannot be evaluated. Not a failure — an absence.
162
- - Mean accuracy 0.6651 against a chance floor of 0.500; 0% of cells fall within 2 points of chance.
163
 
164
  ## Reference models (parents and joint ceilings)
165
 
@@ -202,19 +241,26 @@ Parent B *is* the jointly-trained model, so the joint ceiling is built in and `d
202
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-ita | 0.5344 | 0.4691 | nan | 0.6619 | 0.5593 | 0.6922 | nan | nan | 0.5521 | 0.5128 | 0.6419 | 0.8089 | 0.574 | 0.6027 | 0.5459 | 0.697 | 0.6711 | 0.5231 |
203
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-jpn | 0.5374 | 0.4657 | nan | 0.6214 | 0.552 | 0.6558 | nan | nan | 0.5872 | 0.5294 | 0.6476 | 0.5092 | 0.5272 | 0.6085 | 0.5577 | 0.6793 | 0.6206 | 0.521 |
204
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-kor | 0.5207 | 0.4991 | nan | 0.6449 | 0.5894 | 0.6442 | nan | nan | 0.5342 | 0.5314 | 0.6694 | 0.4872 | 0.5641 | 0.6149 | 0.5371 | 0.6277 | 0.6315 | 0.4971 |
205
- | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-nld | 0.5527 | 0.7346 | nan | 0.6867 | 0.5803 | 0.6779 | nan | nan | 0.5755 | 0.5543 | 0.6566 | 0.5275 | 0.8812 | 0.6143 | 0.494 | 0.6974 | 0.659 | nan |
206
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-pol | 0.5303 | 0.5149 | nan | 0.6371 | 0.6487 | 0.6558 | nan | nan | 0.5739 | 0.5508 | 0.6583 | 0.5405 | 0.4805 | 0.7882 | 0.535 | 0.6781 | 0.6418 | 0.4689 |
207
- | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-rus | 0.5479 | 0.4768 | nan | 0.6079 | 0.6122 | 0.6545 | nan | nan | 0.5728 | 0.4824 | 0.6598 | 0.4538 | 0.4758 | 0.6042 | 0.7941 | nan | nan | nan |
208
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-spa | 0.5495 | 0.4896 | nan | 0.6249 | 0.5931 | 0.7182 | nan | nan | 0.5802 | 0.5128 | 0.6483 | 0.5328 | 0.4895 | 0.6103 | 0.5235 | 0.8977 | 0.6561 | 0.4826 |
209
- | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-tur | 0.522 | 0.4717 | nan | 0.6397 | 0.5383 | 0.6519 | nan | nan | 0.5794 | 0.5218 | 0.6348 | 0.5258 | 0.5307 | 0.6186 | 0.5042 | 0.7135 | 0.7664 | nan |
210
- | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-zho | 0.5303 | 0.4631 | nan | 0.5692 | 0.5347 | 0.6844 | nan | nan | 0.5658 | 0.5135 | 0.6516 | 0.4968 | 0.474 | 0.6027 | nan | nan | nan | nan |
211
  | Beetle-FineWeb-2B/beetle-bilingual-l2-50-sequential-33-67-b3-fineweb-2b-nld-eng | 0.5552 | 0.7683 | nan | nan | nan | 0.7299 | nan | nan | nan | nan | nan | nan | 0.8992 | nan | nan | nan | nan | nan |
 
 
212
  | Beetle-HumanScale/beetle-bilingual-l2-50-classroom-20-b4-humanscale-deu-eng | 0.5702 | nan | nan | 0.8977 | nan | 0.7221 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
 
213
  | Beetle-HumanScale/beetle-bilingual-l2-50-sequential-0-100-b3-humanscale-nld-eng | 0.5739 | 0.7871 | nan | nan | nan | 0.7455 | nan | nan | nan | nan | nan | nan | 0.8975 | nan | nan | nan | nan | nan |
 
 
214
  | Beetle-HumanScale/beetle-bilingual-l2-50-sequential-33-67-b3-humanscale-zho-eng | 0.5472 | nan | nan | nan | nan | 0.7039 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | 0.701 |
215
  | Beetle-HumanScale/beetle-bilingual-l2-50-simultaneous-b2-humanscale-deu-eng | 0.5605 | nan | 43.7057 | 0.8951 | nan | 0.7078 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
216
  | Beetle-HumanScale/beetle-bilingual-l2-50-simultaneous-b2-humanscale-nld-eng | 0.5734 | 0.7914 | 44.776 | nan | nan | 0.7403 | nan | nan | nan | nan | nan | nan | 0.888 | nan | nan | nan | nan | nan |
217
  | Beetle-HumanScale/beetle-bilingual-l2-50-simultaneous-b2-humanscale-zho-eng | 0.5423 | nan | 42.5336 | nan | nan | 0.6883 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | 0.6975 |
 
 
218
  | Beetle-HumanScale/beetle-bilingual-l2-80-late-b5-humanscale-zho-eng | 0.5489 | nan | nan | nan | nan | 0.6961 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | 0.7144 |
219
  | Beetle-HumanScale/beetle-monolingual-humanscale-deu | 0.5115 | 0.4794 | nan | 0.8999 | nan | 0.6468 | nan | nan | nan | nan | nan | nan | 0.4925 | nan | nan | nan | nan | 0.2305 |
220
  | Beetle-HumanScale/beetle-monolingual-humanscale-eng | 0.6472 | 0.492 | 45.7675 | 0.6023 | nan | 0.7896 | nan | nan | nan | nan | nan | nan | 0.5302 | nan | nan | nan | nan | 0.2327 |
 
1
  # Beetle merged models vs their parents — evaluation report
2
 
3
+ _Generated 2026-08-27 02:42 UTC. **PARTIAL RUN** — see the coverage table for exactly what is and is not scored._
4
 
5
 
6
  ## Headline findings
 
48
 
49
  Full audit: `arm_identity_audit.csv`.
50
 
51
+ ## Harness validation
52
+
53
+ The vendored scorer was checked against **beetle-analyze's own published numbers**
54
+ (`results/evals/blimp/blimp_results.csv`) on the **52 (model, benchmark) cells**
55
+ that overlap. Agreement:
56
+
57
+ - median absolute difference **0.00022**; on 67,000-item English BLiMP that is
58
+ a disagreement of about **4 items**.
59
+ - on the 49 cells scoring at or above 0.45, max absolute difference
60
+ **0.01122**.
61
+
62
+ Residual disagreement is expected: the vendored path evaluates the fp32 log-softmax in
63
+ row chunks (`gather - logsumexp`) rather than materialising the whole `[B,T,V]` tensor,
64
+ which reassociates the reduction and flips a handful of near-tied pairs under bf16.
65
+
66
+ **The exception is diagnostic, not cosmetic.** The cells that disagree by ~0.02 are all
67
+ non-Chinese models on ZhoBLiMP, scoring **0.17-0.23 against a chance floor of 0.500** —
68
+ far *below* chance. A model that cannot represent the script produces near-degenerate
69
+ scores where large numbers of pairs are numerically tied, so tie-breaking dominates and
70
+ the number is unstable between runs of the same harness. Those cells should be read as
71
+ "the model has no purchase on this benchmark", never as a measured accuracy. The
72
+ `near_chance` and `z_vs_chance` columns are in the delta CSV for exactly this filtering.
73
+
74
+ Per-cell detail: `harness_validation_vs_beetle_analyze.csv`.
75
+
76
  ## Coverage
77
 
78
  | provenance_family | merges | rungs_published | rungs_scored | complete | partial | not_eval |
79
  |:--------------------|---------:|------------------:|---------------:|-----------:|----------:|-----------:|
80
+ | arch | 1 | 6 | 5 | 0 | 1 | 0 |
81
  | humanscale-pair | 3 | 72 | 72 | 3 | 0 | 0 |
82
+ | zoo-curriculum | 49 | 49 | 8 | 8 | 0 | 41 |
83
+ | zoo-language | 111 | 692 | 52 | 3 | 4 | 104 |
84
+ | zoo-mono_vs_joint | 50 | 50 | 10 | 10 | 0 | 40 |
85
 
86
  Per-merge coverage: `coverage_by_merge.csv`.
87
 
 
89
 
90
  | provenance_family | cells | merges | mean acc | mean Δ vs naive | beats naive | mean Δ vs better parent | beats better parent | mean Δ vs joint ceiling | beats ceiling |
91
  |:--------------------|--------:|---------:|-----------:|------------------:|:--------------|--------------------------:|:----------------------|--------------------------:|:----------------|
92
+ | arch | 20 | 1 | 0.6792 | 0 | 0/16 | -0.0589 | 2/20 | nan | 0/0 |
93
  | humanscale-pair | 240 | 3 | 0.5669 | 0.0055 | 42/240 | -0.1966 | 0/240 | -0.1416 | 4/240 |
94
+ | zoo-curriculum | 24 | 8 | 0.6923 | nan | 0/0 | -0.007 | 6/24 | nan | 0/0 |
95
+ | zoo-language | 208 | 7 | 0.5598 | -0.0003 | 24/171 | -0.1778 | 12/208 | nan | 0/0 |
96
+ | zoo-mono_vs_joint | 29 | 10 | 0.6554 | nan | 0/0 | -0.0787 | 1/29 | -0.0723 | 3/29 |
97
 
98
  ### blimp_eng (chance 0.500)
99
 
100
  | provenance_family | cells | merges | mean acc | mean Δ vs naive | beats naive | mean Δ vs better parent | beats better parent | mean Δ vs joint ceiling | beats ceiling |
101
  |:--------------------|--------:|---------:|-----------:|------------------:|:--------------|--------------------------:|:----------------------|--------------------------:|:----------------|
102
+ | arch | 5 | 1 | 0.5514 | 0 | 0/4 | -0.0038 | 2/5 | nan | 0/0 |
103
  | humanscale-pair | 72 | 3 | 0.5242 | 0.0017 | 14/72 | -0.123 | 0/72 | -0.0345 | 1/72 |
104
+ | zoo-curriculum | 8 | 8 | 0.5538 | nan | 0/0 | -0.0034 | 3/8 | nan | 0/0 |
105
+ | zoo-language | 52 | 7 | 0.5375 | -0.0016 | 5/43 | -0.0459 | 10/52 | nan | 0/0 |
106
+ | zoo-mono_vs_joint | 10 | 10 | 0.54 | nan | 0/0 | -0.048 | 1/10 | -0.0403 | 2/10 |
107
 
108
  ### multiblimp:eng (chance 0.500)
109
 
110
  | provenance_family | cells | merges | mean acc | mean Δ vs naive | beats naive | mean Δ vs better parent | beats better parent | mean Δ vs joint ceiling | beats ceiling |
111
  |:--------------------|--------:|---------:|-----------:|------------------:|:--------------|--------------------------:|:----------------------|--------------------------:|:----------------|
112
+ | arch | 5 | 1 | 0.7151 | 0 | 0/4 | -0.0148 | 0/5 | nan | 0/0 |
113
  | humanscale-pair | 72 | 3 | 0.5609 | 0.0093 | 13/72 | -0.2287 | 0/72 | -0.1512 | 0/72 |
114
+ | zoo-curriculum | 7 | 7 | 0.7022 | nan | 0/0 | -0.0126 | 0/7 | nan | 0/0 |
115
+ | zoo-language | 52 | 7 | 0.6462 | -0.0014 | 5/43 | -0.0689 | 2/52 | nan | 0/0 |
116
+ | zoo-mono_vs_joint | 10 | 10 | 0.666 | nan | 0/0 | -0.1125 | 0/10 | -0.1052 | 1/10 |
117
 
118
  ## Worked example — the three named `beetle-humanscale-*-eng` pairs
119
 
 
147
 
148
  ### arch
149
 
150
+ Scored **5 of 6** published rungs across 1 of 1 repos (20 model-benchmark cells).
151
 
152
  A single cross-architecture pair (pico x MoE).
153
 
154
+ - **vs better parent**: 2/20 cells above; mean Δ -0.0589.
155
  - **vs joint ceiling**: no jointly-trained counterpart published for this family.
156
+ - **vs naive**: 0/16 cells above; mean Δ +0.0000.
157
+ - Mean accuracy 0.6792 against a chance floor of 0.500; 0% of cells fall within 2 points of chance.
158
 
159
  ### humanscale-pair
160
 
 
167
  - **vs naive**: 42/240 cells above; mean Δ +0.0055.
168
  - Mean accuracy 0.5669 against a chance floor of 0.500; 20% of cells fall within 2 points of chance.
169
 
170
+ ### zoo-curriculum
171
+
172
+ Scored **8 of 49** published rungs across 8 of 49 repos (24 model-benchmark cells).
173
+
174
+ Published rungs are `qmd_guided__aligned` only — one rung per repo, no naive and no transport sibling. Δ vs naive is unavailable, and since every aligned folder elsewhere in the org is byte-identical to its naive sibling, these are best read as naive merges.
175
+
176
+ - **vs better parent**: 6/24 cells above; mean Δ -0.0070.
177
+ - **vs joint ceiling**: no jointly-trained counterpart published for this family.
178
+ - **vs naive**: **no naive artefact exists** for these repos, so this bar cannot be evaluated. Not a failure — an absence.
179
+ - Mean accuracy 0.6923 against a chance floor of 0.500; 0% of cells fall within 2 points of chance.
180
+
181
  ### zoo-language
182
 
183
+ Scored **52 of 692** published rungs across 7 of 111 repos (208 model-benchmark cells).
184
 
185
  Largest family. Carries ties/average/task_arithmetic in naive, aligned and transport arms, so it is the only family with broad arm coverage — but the aligned arm is byte-identical to naive throughout.
186
 
187
+ - **vs better parent**: 12/208 cells above; mean Δ -0.1778.
188
  - **vs joint ceiling**: no jointly-trained counterpart published for this family.
189
+ - **vs naive**: 24/171 cells above; mean Δ -0.0003.
190
+ - Mean accuracy 0.5598 against a chance floor of 0.500; 21% of cells fall within 2 points of chance.
191
 
192
  ### zoo-mono_vs_joint
193
 
194
+ Scored **10 of 50** published rungs across 10 of 50 repos (29 model-benchmark cells).
195
 
196
  Parent B *is* the jointly-trained model, so the joint ceiling is built in and `d_ceiling` is meaningful here. Published rungs are `qmd_guided__aligned` and `qmd_guided__transport` only — no naive folder exists, so Δ vs naive is unavailable by construction.
197
 
198
+ - **vs better parent**: 1/29 cells above; mean Δ -0.0787.
199
+ - **vs joint ceiling**: 3/29 cells above; mean Δ -0.0723.
200
  - **vs naive**: **no naive artefact exists** for these repos, so this bar cannot be evaluated. Not a failure — an absence.
201
+ - Mean accuracy 0.6554 against a chance floor of 0.500; 7% of cells fall within 2 points of chance.
202
 
203
  ## Reference models (parents and joint ceilings)
204
 
 
241
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-ita | 0.5344 | 0.4691 | nan | 0.6619 | 0.5593 | 0.6922 | nan | nan | 0.5521 | 0.5128 | 0.6419 | 0.8089 | 0.574 | 0.6027 | 0.5459 | 0.697 | 0.6711 | 0.5231 |
242
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-jpn | 0.5374 | 0.4657 | nan | 0.6214 | 0.552 | 0.6558 | nan | nan | 0.5872 | 0.5294 | 0.6476 | 0.5092 | 0.5272 | 0.6085 | 0.5577 | 0.6793 | 0.6206 | 0.521 |
243
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-kor | 0.5207 | 0.4991 | nan | 0.6449 | 0.5894 | 0.6442 | nan | nan | 0.5342 | 0.5314 | 0.6694 | 0.4872 | 0.5641 | 0.6149 | 0.5371 | 0.6277 | 0.6315 | 0.4971 |
244
+ | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-nld | 0.5527 | 0.7346 | nan | 0.6867 | 0.5803 | 0.6779 | nan | nan | 0.5755 | 0.5543 | 0.6566 | 0.5275 | 0.8812 | 0.6143 | 0.494 | 0.6974 | 0.659 | 0.4956 |
245
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-pol | 0.5303 | 0.5149 | nan | 0.6371 | 0.6487 | 0.6558 | nan | nan | 0.5739 | 0.5508 | 0.6583 | 0.5405 | 0.4805 | 0.7882 | 0.535 | 0.6781 | 0.6418 | 0.4689 |
246
+ | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-rus | 0.5479 | 0.4768 | nan | 0.6079 | 0.6122 | 0.6545 | nan | nan | 0.5728 | 0.4824 | 0.6598 | 0.4538 | 0.4758 | 0.6042 | 0.7941 | 0.6907 | 0.6487 | 0.5077 |
247
  | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-spa | 0.5495 | 0.4896 | nan | 0.6249 | 0.5931 | 0.7182 | nan | nan | 0.5802 | 0.5128 | 0.6483 | 0.5328 | 0.4895 | 0.6103 | 0.5235 | 0.8977 | 0.6561 | 0.4826 |
248
+ | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-tur | 0.522 | 0.4717 | nan | 0.6397 | 0.5383 | 0.6519 | nan | nan | 0.5794 | 0.5218 | 0.6348 | 0.5258 | 0.5307 | 0.6186 | 0.5042 | 0.7135 | 0.7664 | 0.4747 |
249
+ | Beetle-FineWeb-100M/beetle-monolingual-fineweb-100m-zho | 0.5303 | 0.4631 | nan | 0.5692 | 0.5347 | 0.6844 | nan | nan | 0.5658 | 0.5135 | 0.6516 | 0.4968 | 0.474 | 0.6027 | 0.5188 | 0.6812 | 0.6688 | 0.6458 |
250
  | Beetle-FineWeb-2B/beetle-bilingual-l2-50-sequential-33-67-b3-fineweb-2b-nld-eng | 0.5552 | 0.7683 | nan | nan | nan | 0.7299 | nan | nan | nan | nan | nan | nan | 0.8992 | nan | nan | nan | nan | nan |
251
+ | Beetle-HumanScale/beetle-bilingual-balanced-b1-humanscale-deu-eng | 0.6546 | nan | nan | 0.869 | nan | 0.8052 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
252
+ | Beetle-HumanScale/beetle-bilingual-balanced-b1-humanscale-nld-eng | 0.6906 | 0.7687 | nan | nan | nan | 0.8078 | nan | nan | nan | nan | nan | nan | 0.8752 | nan | nan | nan | nan | nan |
253
  | Beetle-HumanScale/beetle-bilingual-l2-50-classroom-20-b4-humanscale-deu-eng | 0.5702 | nan | nan | 0.8977 | nan | 0.7221 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
254
+ | Beetle-HumanScale/beetle-bilingual-l2-50-classroom-20-b4-humanscale-nld-eng | 0.5841 | 0.7871 | nan | nan | nan | 0.7403 | nan | nan | nan | nan | nan | nan | 0.8983 | nan | nan | nan | nan | nan |
255
  | Beetle-HumanScale/beetle-bilingual-l2-50-sequential-0-100-b3-humanscale-nld-eng | 0.5739 | 0.7871 | nan | nan | nan | 0.7455 | nan | nan | nan | nan | nan | nan | 0.8975 | nan | nan | nan | nan | nan |
256
+ | Beetle-HumanScale/beetle-bilingual-l2-50-sequential-33-67-b3-humanscale-deu-eng | 0.5743 | nan | nan | 0.8995 | nan | 0.713 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
257
+ | Beetle-HumanScale/beetle-bilingual-l2-50-sequential-33-67-b3-humanscale-nld-eng | 0.5825 | 0.7837 | nan | nan | nan | 0.7377 | nan | nan | nan | nan | nan | nan | 0.9026 | nan | nan | nan | nan | nan |
258
  | Beetle-HumanScale/beetle-bilingual-l2-50-sequential-33-67-b3-humanscale-zho-eng | 0.5472 | nan | nan | nan | nan | 0.7039 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | 0.701 |
259
  | Beetle-HumanScale/beetle-bilingual-l2-50-simultaneous-b2-humanscale-deu-eng | 0.5605 | nan | 43.7057 | 0.8951 | nan | 0.7078 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
260
  | Beetle-HumanScale/beetle-bilingual-l2-50-simultaneous-b2-humanscale-nld-eng | 0.5734 | 0.7914 | 44.776 | nan | nan | 0.7403 | nan | nan | nan | nan | nan | nan | 0.888 | nan | nan | nan | nan | nan |
261
  | Beetle-HumanScale/beetle-bilingual-l2-50-simultaneous-b2-humanscale-zho-eng | 0.5423 | nan | 42.5336 | nan | nan | 0.6883 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | 0.6975 |
262
+ | Beetle-HumanScale/beetle-bilingual-l2-80-late-b5-humanscale-deu-eng | 0.5579 | nan | nan | 0.8956 | nan | 0.7117 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |
263
+ | Beetle-HumanScale/beetle-bilingual-l2-80-late-b5-humanscale-nld-eng | 0.5664 | 0.7942 | nan | nan | nan | 0.7247 | nan | nan | nan | nan | nan | nan | 0.8846 | nan | nan | nan | nan | nan |
264
  | Beetle-HumanScale/beetle-bilingual-l2-80-late-b5-humanscale-zho-eng | 0.5489 | nan | nan | nan | nan | 0.6961 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | 0.7144 |
265
  | Beetle-HumanScale/beetle-monolingual-humanscale-deu | 0.5115 | 0.4794 | nan | 0.8999 | nan | 0.6468 | nan | nan | nan | nan | nan | nan | 0.4925 | nan | nan | nan | nan | 0.2305 |
266
  | Beetle-HumanScale/beetle-monolingual-humanscale-eng | 0.6472 | 0.492 | 45.7675 | 0.6023 | nan | 0.7896 | nan | nan | nan | nan | nan | nan | 0.5302 | nan | nan | nan | nan | 0.2327 |