Hugging Face's logo

Datasets:
egolimblevskaia
/
circuitlens-gemma-2-2btranscoder-descriptions-and-evaluations

Tasks:
Text Classification
Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
interpretability
mechanistic-interpretability
transcoders
sparse-autoencoders
gemma-2
Libraries:
Datasets
Dask
Polars
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
neuron_idx
int64
1
16.4k
description
stringlengths
4
683
⌀
Clarity
float64
0
1
⌀
Responsiveness
float64
0
1
Purity
float64
0
1
Faithfulness
float64
0
1
⌀
24
Frequent use of the phrase “is true” and variations, often in contexts of expectation or observation.
0.733513
0.291869
0.015597
0.04
58
Frequent use of "et al" in academic references, often following names or terms, indicating multiple authors or contributors.
0.990464
0.512021
0.105286
0
106
Repeated use of the token "cc" and various LaTeX formatting commands, indicating mathematical or technical notation.
0.544596
0.128487
0.12577
0.020408
130
Statements indicating causation or implications, often using verbs like “implies”, “requires”, or “concludes”.
0.344988
0.03007
0.106074
0
150
Frequent references to "system" or "systems," often preceded by various descriptors or phrases (e.g., "the system," "nonlinear system").
0.836654
0.884813
0.58528
0.02
176
Repeated references to the variable 'Z' and terms related to mathematical concepts, including 'negative' and 'events'.
0.871883
0.116335
0.07747
0.02
184
LaTeX formatting for mathematical expressions, encompassing commands for operators, subscripts, superscripts, symbols, and various brackets.
0.000309
0.858078
0.870709
0.02
201
Various scientific notation and formatting elements, including units (e.g., mm, MeV), LaTeX commands, and references to figures or tables.
0.803946
0.860868
0.863244
0.02
222
Repeated phrases indicating potential outcomes or observations, particularly “can result in” and “other hand”.
0.942006
0.324881
0.037652
0
260
Repeated use of the term "prime" with various mathematical notations, particularly superscripts and braces.
0.142165
0.378874
0.051783
0
323
Mathematical notation and formatting elements, including symbols, cases, and LaTeX commands, often involving variables and expressions.
0.134519
0.452096
0.275232
0.041667
324
Repetitive structures using phrases like “it is clear” and “we find” indicating conclusions or observations.
0.47423
0.195041
0.02745
0
328
Repeated use of the LaTeX command “\operatorname” in mathematical or scientific contexts, often followed by variables or terms.
null
0.017589
0.026939
0
336
Repetitive phrases indicating actions or observations, particularly “carried out”, “pointed out”, and “turns out”.
0.949519
0.55101
0.038983
0
363
Repeated references to "features" and various types of "graphs" in a scientific or mathematical context.
0.911386
0.147941
0.051575
0
378
Repeated use of "et al" for citations, often followed by various phrases indicating research findings or characteristics.
0.967533
0.671367
0.131323
0
417
Frequent use of "potentially" and "community" in various contexts, often related to policies or theories.
0.985593
0.945383
0.567612
0
429
Symbols and formatting elements like '&', '(', and '\n' in a structured layout, often indicating mathematical or programming contexts.
0.157167
0.420338
0.368975
0.12766
433
Frequent mentions of "cross section" and "half life" in scientific contexts, often accompanied by HTML-like span tags for formatting.
0.997372
0.760649
0.008403
0
459
Repeated references to formulas, universities, and various methods, often preceded by “the” or “of”.
0.232796
0.074305
0.18319
0.122449
487
Frequent use of "subject", "suppose", and "full" in contexts related to mathematical or scientific terms.
0.776095
0.220419
0.166371
0
505
References to URLs or HTTP links, often accompanied by additional context or qualifiers (e.g., "http [...]").
0.312118
0.028618
0.026225
0
512
Mathematical expressions and symbols, including equations, functions, and LaTeX formatting, often involving parentheses and equal signs.
0.067476
0.407346
0.28718
0.04
539
Repeated use of mathematical symbols and formatting, particularly involving "colour", "sqrt", and "rightarrow" in LaTeX notation.
null
0.467991
0.102067
0
551
Frequent use of "associated" and "action" in various contexts, often followed by specific terms or phrases.
0.995139
0.975359
0.93067
0.02
556
Repeated instances of the token ‘a’ followed by a dollar sign ($), indicating mathematical or financial notation.
0.386249
0.360965
0.1627
0.040816
567
A variety of scientific terms and concepts, including mathematical symbols, physical theories, and technical jargon across multiple disciplines.
0.343406
0.767961
0.834376
0.020833
575
Frequent use of present perfect and past participle forms, particularly with “have” or “has” followed by verbs (e.g., “have shown”, “has been used”).
0.907089
0.934893
0.873882
0.021739
620
Frequent use of the phrase "the" followed by various nouns, indicating a pattern of specification or reference.
0.392418
0.359113
0.30874
0.162791
649
Repeated references to mathematical or scientific terms followed by “quantities”, “index”, or “open”, alongside potential actions indicated by “can” followed by verbs.
0.207319
0.235284
0.214005
0
671
Frequent use of passive constructions with forms of "be" (e.g., "is", "be", "been") indicating actions or states.
0.008405
0.447603
0.387929
0.021277
702
Repeated use of the terms "disc" or "disk" and "basis" in various contexts, often followed by additional descriptors or qualifiers.
0.98797
0.983047
0.881594
0
749
Frequent use of negation phrases like "not", "no", "un-", and "unless" indicating exceptions or limitations.
0.766653
0.880192
0.767716
0.061224
757
Frequent use of "multi" and "cross section" in various contexts, often indicating multi-dimensional or multi-faceted concepts.
0.458995
0.282398
0.045323
0
766
Frequent references to studies, observations, and combinations, often prefixed by “this” or “one” (e.g., "this study", "one such").
0.907536
0.038586
0.14424
0
767
References to authors (e.g., Mar, Sch) and repeated phrases like “are handled” and “focusing”.
0.960134
0.131831
0.086278
0
768
Repeated use of LaTeX formatting commands and symbols, including brackets, dollar signs, and various mathematical notations.
0.447721
0.07466
0.101787
0.02
774
Various punctuation marks and formatting symbols, including colons, braces, and line breaks, often in mathematical or programming contexts.
0.519323
0.843508
0.882452
0.093023
779
Frequent references to the variable 'z' in mathematical or scientific contexts, often accompanied by other variables or symbols.
0.065161
0.270145
0.106434
0.02
790
Repeated use of the suffix “-ness” and phrases indicating construction or modification (e.g., "to construct", "settles down").
0.214122
0.03957
0.009297
0
797
Tokens primarily consist of Basque language elements, including repeated phrases and names, alongside mathematical or formal symbols.
0.838118
0.37345
0.052729
0
811
Repeated phrases involving "average" and "regime", often in scientific or mathematical contexts (e.g., "on average", "static regime").
0.674342
0.110292
0.024473
0
822
Mathematical definitions and expressions, often using LaTeX formatting with symbols like “defined”, “frac”, and “mathrm”.
null
0.321102
0.178678
0
830
Repeated phrases involving "the whole" and "the point," indicating specific references or locations in a context.
0.989064
0.753181
0.077295
0.061224
878
Frequent use of modal verbs (will, may, can, must) followed by "be" in passive constructions.
0.927403
0.710446
0.458445
0
892
LaTeX formatting with repeated use of subscripts, superscripts, and specific commands like \textit and \mathrm.
0.0165
0.056416
0.069168
0
923
Repeated use of scientific or technical terms, often with specific suffixes like “-ates”, “-al”, or “-ic” (e.g., compartmental, asymptote).
0.420007
0.650866
0.726539
0.020408
979
Repeated references to "area", "environment", and "ion", often preceded by articles like "the" or "an".
0.96982
0.672683
0.135043
0
1,083
Repeated use of the term "note" in various contexts, often followed by mathematical or scientific notation.
0.408858
0.069742
0.040904
0
1,111
Repeated use of terms related to generation or relation, often followed by “by” or “relation” (e.g., "generated by", "relation").
0.955011
0.567677
0.397207
0.04
1,158
Repeated use of “remains” and “above” to reference previously mentioned concepts or data in a structured format.
0.985853
0.671128
0.067025
0.06
1,174
Frequent use of conjunctions like "and", "but", and "that" to connect clauses in complex sentences.
0.161442
0.482515
0.672232
0.065217
1,181
Mathematical equations or expressions consistently formatted with an equals sign '=' followed by a value or expression.
0.065757
0.236658
0.071526
0.06
1,182
Repeated use of the term "effective" and "free" in a scientific or mathematical context, often followed by specific terms or symbols.
0.953218
0.492708
0.058181
0
1,203
Repeated use of "is/are" with adjectives like "achieved", "connected", "exactly", and "precisely" in formal contexts.
0.847574
0.266267
0.09552
0.02
1,273
Conditional and comparative phrases, frequently using “if”, “when”, “as well”, and “for example”.
0.215407
0.201213
0.121934
0.08
1,325
Repeated use of "bound" in various contexts, often preceded by qualifiers like "upper" or "lower".
0.924162
0.899625
0.513193
0
1,384
Phrases like "as seen" and "well understood" in formal contexts, alongside repeated references to online sources such as arXiv, often formatted as "on line" or "arXiv astro".
0.636401
0.172395
0.041349
0.02
1,424
Repeated references to Greek letters, especially 'alpha', 'beta', and 'gamma', in mathematical or scientific contexts.
0.310531
0.060028
0.035977
0
1,457
Repeated phrases involving “treated” and “hypothesis”, along with various system and transform references, often followed by “the” or “that”.
0.989316
0.9695
0.902351
0.02
1,482
Abbreviated terms and phrases, likely representing titles, sections, or concepts in academic or technical contexts.
0.032058
0.428984
0.296248
0.021277
1,503
Technical terms and phrases often ending with nouns, indicating conditions, systems, or properties (e.g., "initial conditions", "energy gap").
0.340034
0.268816
0.278659
0.06
1,532
Frequent references to "single" and "control" in various contexts, often associated with collaborations or methodologies.
0.863434
0.536363
0.082042
0
1,550
LaTeX mathematical expressions and equations, frequently using the underscore for subscripts and dollar signs for inline math.
null
0.937044
0.935296
0
1,602
Adjectives and adverbial phrases indicating degree or extent, often followed by comparative or superlative structures (e.g., "more clearly", "larger fraction").
0.801091
0.634822
0.349941
0.06
1,610
Technical terms and phrases from physics, engineering, and data science, often featuring 'field', 'energy', and 'magnetic' in various contexts.
0.877931
0.522968
0.327946
0
1,692
References to various types of systems (e.g., "mechanical system", "planetary systems") and studies or papers.
0.923487
0.492793
0.227056
0
1,719
Frequent occurrences of abbreviations like "Phys", "e", and "eps" in a structured format, often followed by punctuation or symbols.
0.945624
0.321736
0.154607
0.02
1,729
Frequent references to "light" and "property" in various contexts, often with modifiers like "reduced" or "negligible".
0.883024
0.580473
0.050336
0
1,747
Mathematical expressions and variables, particularly involving 'a', 'b', and 'i', often formatted in LaTeX with various symbols and brackets.
0.865491
0.674492
0.380538
0.02
1,748
Mathematical expressions and structures involving tuples, vectors, parameters, and classes, often formatted with LaTeX symbols like $(...)$ and $\{...\}$.
null
0.620546
0.740333
0
1,799
LaTeX formatting for mathematical expressions, utilizing commands like \mathbf, \mathcal, subscripts, and symbols, often with braces and underscores.
0.110837
0.793178
0.742559
0.02
1,821
Frequent references to "level" and "characteristic" in various contexts, often with modifiers like "near" or "full".
0.90705
0.450283
0.036317
0.02
1,846
Repeated phrases starting with "to" followed by verbs indicating actions or processes (e.g., "to measure", "to estimate").
0.55546
0.380816
0.135553
0.061224
1,898
Repeated references to the symbol "C" and its variations, often in mathematical contexts, alongside terms like "scale" and "mirror".
0.259398
0.2473
0.044292
0.02
1,935
Usage of "various" and "additional" to indicate multiple or extra elements in mathematical or scientific contexts.
0.966585
0.566962
0.249088
0.02
1,940
Repeated references to “site” or “sites” in various contexts, often with modifiers or descriptors.
0.183578
0.98353
0.942711
0
2,133
Frequent use of the word "large" in various contexts, often describing size or magnitude.
0.981193
0.685513
0.105739
0
2,171
Repeated use of terms like “determined” and “namely” in a scientific or mathematical context, often followed by specific variables or objects.
0.893572
0.227426
0.1303
0
2,195
Frequent use of "the" followed by various nouns, indicating specific references or topics in a structured context.
0.471183
0.331986
0.359389
0.119048
2,219
References to various academic or scientific notations, including citations, mathematical symbols, and special characters.
0.04542
0.107235
0.243656
0
2,223
Frequent use of "the" and "a" with phrases indicating relationships or comparisons (e.g., "the observed", "similar to").
0.28384
0.370024
0.414065
0.391304
2,245
Repeated use of mathematical terms and structures, including “sets”, “categories”, and “graphs” in formal contexts.
0.657497
0.199882
0.149601
0
2,255
Frequent use of "holds" and "near" to indicate conditions or explanations, alongside the phrase "with respect" to establish a consistent reference point.
0.977708
0.734612
0.272028
0
2,258
References to mathematical functions, procedures, and types, often using terms like “function”, “procedure”, and “quotient”.
0.798199
0.6812
0.504693
0.081633
2,265
Frequent mentions of "kernel" in various contexts, often combined with descriptors like "heat", "filter", or "smoothing".
0.191253
0.406439
0.003378
0
2,266
Repeated use of LaTeX commands, particularly “\left” and “\quad”, in mathematical or formal contexts.
0.598262
0.027574
0.067796
0
2,291
Mathematical symbols and formatting, particularly superscripts, subscripts, and various operators, often in LaTeX notation.
0.29722
0.118735
0.131962
0.020408
2,313
Frequent use of "I" in personal statements, often followed by verbs or phrases indicating opinion or results (e.g., "I think", "I model").
0.422401
0.80504
0.617493
0.08
2,317
Frequent use of negation with terms like "not", "no", and "non" to express limitations or exclusions.
0.813376
0.856835
0.710175
0
2,356
Frequent use of “that” and “which” to introduce clauses, often followed by verbs or phrases indicating relationships or characteristics.
0.584433
0.74602
0.73113
0.020833
2,449
Mathematical expressions and equations featuring variables, symbols, and formatting, often including inequalities and equalities.
0.628262
0.350536
0.232023
0.02
2,454
Repeated references to 'alpha' and 'beta' as variables or parameters in mathematical or scientific contexts.
0.170277
0.077257
0.054007
0
2,472
Frequent references to "the presence" and "is reduced" in various contexts, often indicating conditions or results.
0.986293
0.92891
0.635333
0.02
2,475
Various scientific or mathematical terms and concepts, often paired with adjectives indicating types or classifications (e.g., "modal Australian", "quadratic Gaussian").
0.10715
0.38941
0.221702
0
2,522
Repeated use of "without" in various contexts, often indicating conditions or exceptions.
0.469731
0.466005
0.01788
0
2,541
Technical terms and symbols in scientific or mathematical contexts, including references to equations, thresholds, and bounded conditions.
0.216488
0.012673
0.230427
0.022222
2,566
Repetitions of the term "metallicity" and phrases involving "to" or "and" indicating connections or comparisons.
0.088467
0.321094
0.030059
0
2,642
LaTeX mathematical formatting featuring repeated brackets, equations, and symbols to denote mathematical expressions or conditions.
0.461556
0.706827
0.621333
0
2,655
Frequent passive constructions using “is considered” and “are considered” related to mass and flow, indicating relationships or comparisons in various contexts.
0.969824
0.389456
0.094026
0
End of preview. Expand in Data Studio

CircuitLens & WeightLens: Transcoder Descriptions and Evaluations

This dataset contains automatically generated descriptions and evaluation metrics for Gemma-2-2B transcoders, produced using CircuitLens and WeightLens methods.

Methods

Dataset Structure

The dataset is organized by layers (0, 4, 7, 10, 12, 15, 18, 21, 23, 25), with each layer containing CSV files for different description methods:

├── layer_0/
│   ├── WeightLens.csv
│   ├── WeightLens_Out.csv
│   ├── WeightLens_Out_LLM.csv
│   ├── CircuitLens-Input.csv
│   ├── CircuitLens-Full.csv
│   ├── CircuitLens-Input_top.csv
│   ├── CircuitLens-Full_top.csv
│   ├── WL_CircuitLens-Full.csv
│   ├── Neuronpedia.csv
│   └── MaxAct.csv
├── layer_4/
│   └── ...
└── ...

CSV Columns

Column Description
neuron_idx Transcoder neuron index
description Generated concept description
Clarity Clarity evaluation metric
Responsiveness Responsiveness evaluation metric
Purity Purity evaluation metric
Faithfulness Faithfulness evaluation metric

Paper

Circuit Insights: Towards Interpretability Beyond Activations

Elena Golimblevskaia, Aakriti Jain, Bruno Puri, Ammar Ibrahim, Wojciech Samek, and Sebastian Lapuschkin

https://arxiv.org/abs/2510.14936

Citation 

@article{golimblevskaia2024circuit,
  title={Circuit Insights: Towards Interpretability Beyond Activations},
  author={Golimblevskaia, Elena and Jain, Aakriti and Puri, Bruno and Ibrahim, Ammar and Samek, Wojciech and Lapuschkin, Sebastian},
  journal={arXiv preprint arXiv:2510.14936},
  year={2024}
}
Downloads last month
553
Number of rows:
48,551
Total file size:
9.39 MB

Paper for egolimblevskaia/circuitlens-gemma-2-2btranscoder-descriptions-and-evaluations