model_id
stringlengths
19
121
model_role
stringclasses
5 values
provenance_family
stringclasses
35 values
parents
stringclasses
328 values
rung
stringclasses
108 values
operator
stringclasses
52 values
experiment
stringclasses
17 values
benchmark
stringclasses
84 values
metric
stringclasses
13 values
value
float64
-5.09
348
chance
float64
0
50
n
float64
273
67k
source_dataset
stringclasses
7 values
notes
stringclasses
92 values
metric_kind
stringclasses
6 values
unit
stringclasses
7 values
mergebench:Llama-3.1-8B-Instruct|RegMean
merged
Llama-3.1-8B-Instruct
Llama-3.1-8B-Instruct: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
89.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 17); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B-Instruct|TIES
merged
Llama-3.1-8B-Instruct
Llama-3.1-8B-Instruct: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
90.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 17); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B-Instruct|Task arithmetic
merged
Llama-3.1-8B-Instruct
Llama-3.1-8B-Instruct: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
79.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 17); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|Consensus TA
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
60.7
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|DARE
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
55.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|Dataless L&S
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
79.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|Fisher Merging
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
76
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|L&S
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
63.7
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|Model soup
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
71
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|RegMean
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
67.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|TIES
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
61.9
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.1-8B|Task arithmetic
merged
Llama-3.1-8B
Llama-3.1-8B: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
59
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 16); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|Consensus TA
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
80
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|DARE
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
94
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|Dataless L&S
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
85.1
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|Fisher Merging
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
85.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|L&S
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
83.1
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|Model soup
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
84.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|RegMean
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
87.5
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|TIES
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
94.5
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B-Instruct|Task arithmetic
merged
Llama-3.2-3B-Instruct
Llama-3.2-3B-Instruct: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
80.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 13); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|Consensus TA
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
48.2
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|DARE
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
44.9
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|Dataless L&S
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
37.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|Fisher Merging
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
35.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|L&S
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
41.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|Model soup
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
39.2
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|RegMean
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
71.7
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|TIES
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
40.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:Llama-3.2-3B|Task arithmetic
merged
Llama-3.2-3B
Llama-3.2-3B: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
46.1
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 12); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|Consensus TA
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
81
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|DARE
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
74.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|Dataless L&S
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
79.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|Fisher Merging
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
76.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|L&S
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
79
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|Model soup
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
81.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|RegMean
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
83.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|TIES
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
76.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b-it|Task arithmetic
merged
gemma-2-2b-it
gemma-2-2b-it: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
81.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 11); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|Consensus TA
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
61.9
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|DARE
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
39.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|Dataless L&S
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
65
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|Fisher Merging
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
58.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|L&S
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
61.9
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|Model soup
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
52.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|RegMean
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
56.8
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|TIES
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
38.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-2b|Task arithmetic
merged
gemma-2-2b
gemma-2-2b: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
45.1
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 10); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|Consensus TA
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
77.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|DARE
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
62.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|Dataless L&S
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
79
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|Fisher Merging
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
74.7
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|L&S
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
69.7
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|Model soup
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
68.2
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|RegMean
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
73.1
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|TIES
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
71.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b-it|Task arithmetic
merged
gemma-2-9b-it
gemma-2-9b-it: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
75.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 15); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|Consensus TA
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
Consensus TA
mergebench_published_outcomes
Safety
benchmark_score_pct
72.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|DARE
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
DARE
mergebench_published_outcomes
Safety
benchmark_score_pct
73.2
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|Dataless L&S
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
Dataless L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
68.1
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|Fisher Merging
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
Fisher Merging
mergebench_published_outcomes
Safety
benchmark_score_pct
81.7
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|L&S
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
L&S
mergebench_published_outcomes
Safety
benchmark_score_pct
72.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|Model soup
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
Model soup
mergebench_published_outcomes
Safety
benchmark_score_pct
70.6
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|RegMean
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
RegMean
mergebench_published_outcomes
Safety
benchmark_score_pct
84.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|TIES
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
TIES
mergebench_published_outcomes
Safety
benchmark_score_pct
75.3
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
mergebench:gemma-2-9b|Task arithmetic
merged
gemma-2-9b
gemma-2-9b: all 5 domain experts
5-expert merge
Task arithmetic
mergebench_published_outcomes
Safety
benchmark_score_pct
74.4
null
null
Mergeability-2/mergebench-property-sweep
published by MergeBench (arXiv:2505.10833 Table 14); every released score is a 5-expert merge, so no pair-level outcome exists. Suite aggregate over heterogeneous tasks -- NOT a single-benchmark accuracy, so no chance level is defined.
benchmark_score
percent_0_100
RECON:en_el:100:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=100.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.130829
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:10:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=10.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.221846
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:25:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=25.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.074629
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:33:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=33.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.019508
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
8.913833
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:100:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=100.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
8.905946
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:10:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=10.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.046599
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:10:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=10.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
10.050169
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:20:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=20.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.602
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:20:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=20.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
9.600392
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:25:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=25.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.456176
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:25:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=25.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
9.452557
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:33:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=33.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.307967
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:33:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=33.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
9.303435
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:50:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=50.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.085877
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:50:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=50.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
9.078075
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:66:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=66.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
8.951383
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:66:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=66.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
8.943798
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:75:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=75.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
8.926729
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:50:75:trained
jointly_trained
Goldfish en_el
null
alpha=50.0,topk=75.0
null
goldfish_crosslingual
held_out[el]
nats_per_token
8.919005
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:66:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=66.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
8.918031
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:75:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=75.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
8.940999
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
RECON:en_el:90:100:merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=90.0,topk=100.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
9.0227
null
500
suchirsalhan/goldfish-crosslingual-cka
config=reconstructed; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha10-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=10.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha100-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=100.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha25-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=25.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha33-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=33.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha50-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=50.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha66-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=66.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha75-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=75.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el-alpha90-TOKENWISE_MERGE
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=90.0,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.315835
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el_merge-merged
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=,topk=
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.315835
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el_merge-trained
jointly_trained
Goldfish en_el
null
alpha=,topk=
null
goldfish_crosslingual
held_out[el]
nats_per_token
10.312775
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=trained. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. 'trained' = a model trained directly on the mixture, the merge's comparator. | LIKELIHOOD, not accuracy
likelihood
nats/token
suchirsalhan/en_el_merge_top10-merged-top10
merged
Goldfish en_el
goldfish[en]|goldfish[el]
alpha=,topk=10.0
merged
goldfish_crosslingual
held_out[el]
nats_per_token
10.33516
null
500
suchirsalhan/goldfish-crosslingual-cka
config=as-uploaded; arm=merged. NLL only -- the CKA columns in the source are an intrinsic representation-similarity diagnostic and are not extrinsic evaluations. | LIKELIHOOD, not accuracy
likelihood
nats/token