Datasets:

BurnyCoder
/

length-penalized-autointerp-gpt2-small

latent int64 56 24.5k	condition stringclasses 6 values	budget int64 2 40 ⌀	explanation stringlengths 8 291	parsed bool 1 class
56	gen_b2	2	Word fragments	true
71	gen_b2	2	Idiomatic phrase	true
56	gen_b5	5	Suffixes forming English words.	true
594	gen_b5	5	Proper noun "Phoenix" consistently activated.	true
594	gen_b2	2	City name	true
71	gen_b5	5	Phrases indicating a long distance.	true
56	gen_b10	10	Suffixes and word endings in various contexts and domains.	true
594	gen_b10	10	Proper noun "Phoenix" consistently activated across various contexts.	true
71	gen_b10	10	Idiomatic phrase "a long way" indicating progress or distance.	true
71	gen_b20	20	The phrase "a long way" is used to convey progress, distance, or time in various contexts.	true
56	gen_b20	20	Suffixes and word endings, often indicating grammatical function or word type, such as verb or noun forms.	true
594	gen_b20	20	Proper noun "Phoenix" consistently activated, often referring to the city in Arizona or a mythological bird.	true
56	gen_unconstrained	null	Suffixes or word endings, often indicating a relationship or a state, such as possession, location, or action.	true
71	gen_unconstrained	null	The phrase "a long way" is used to convey a significant amount of progress, time, or distance, often in the context of improvement, change, or journey.	true
679	gen_b2	2	Proper nouns	true
638	gen_b2	2	Prepositions	true
638	gen_b5	5	Prepositions indicating location or direction.	true
56	gen_b40	40	Tokens often represent suffixes or word endings, typically indicating a change in word form or function, such as comparative or superlative forms, or verb conjugations, and sometimes denoting a specific part of speech or grammatical function.	true
71	gen_b40	40	The phrase "a long way" is consistently used to convey a significant amount of progress, time, or distance, often in the context of improvement, change, or journey, and is frequently used in idiomatic expressions.	true
638	gen_b10	10	Prepositions indicating location, direction, or relation to something.	true
679	gen_b5	5	Names and brand titles.	true
638	gen_b20	20	Prepositions indicating location, direction, or relationship, often used in phrases describing movement or position.	true
780	gen_b5	5	Geographic location referred to city.	true
780	gen_b2	2	Urban location	true
638	gen_unconstrained	null	Prepositions indicating location, direction, or relationship, often used in phrases that describe movement, position, or interaction.	true
638	gen_b40	40	Prepositions indicating location, direction, or relationship, often used to describe movement, position, or interaction, frequently appearing in phrases that convey a sense of physical or abstract connection between entities or concepts.	true
780	gen_b10	10	Geographic location or urban area referred to as "city".	true
679	gen_b10	10	Names of organizations, titles, and proper nouns with suffixes.	true
679	gen_b20	20	Proper nouns, often names of organizations, people, or titles, with distinctive formatting and punctuation.	true
594	gen_b40	40	The word "Phoenix" is consistently activated across various contexts, including news articles, event descriptions, and fictional narratives, often referring to the city of Phoenix, Arizona, or the mythological bird, indicating a proper noun or title.	true
679	gen_unconstrained	null	Proper nouns, often names of organizations, people, or titles, sometimes with special characters or punctuation, such as hyphens or exclamation marks.	true
594	gen_unconstrained	null	The proper noun "Phoenix" consistently appears as a significant token, often referring to the city of Phoenix, Arizona, but also used in other contexts such as a name, a title, or a concept, indicating a high level of importance across various domains.	true
1,037	gen_b2	2	Nouns categorization	true
679	gen_b40	40	Proper nouns, often names of organizations, people, or titles, with distinctive formatting, such as hyphenation or unusual capitalization, and sometimes accompanied by punctuation marks, like exclamation points or quotation marks.	true
1,037	gen_b5	5	Nouns representing concepts and categories.	true
1,037	gen_b10	10	Nouns and adjectives representing categories and concepts.	true
1,093	gen_b2	2	Preposition "at"	true
780	gen_unconstrained	null	The word "city" consistently appears in contexts where it refers to a specific, often named, urban location, and is frequently associated with descriptions of its characteristics, history, or geography.	true
780	gen_b20	20	The word "city" is consistently highlighted, indicating its importance in various contexts, often referring to a specific location.	true
1,093	gen_b5	5	Preposition indicating a location.	true
1,037	gen_b20	20	Nouns and adjectives representing various categories, concepts, and attributes, often related to identity, characteristics, and activities.	true
1,093	gen_b20	20	Preposition "at" indicating a location or venue for performances, events, or activities.	true
1,093	gen_b10	10	Preposition "at" indicating location or venue of an event.	true
1,269	gen_b5	5	"Major" in professional sports contexts.	true
780	gen_b40	40	The word "city" is consistently highlighted across various contexts, suggesting its importance in describing urban areas, often in relation to history, culture, geography, and spirituality, and is frequently used as a noun to refer to a distinct entity.	true
1,269	gen_b2	2	Major League	true
1,269	gen_b10	10	"Major" referring to top-level professional sports leagues.	true
1,037	gen_unconstrained	null	Nouns and adjectives representing categories, characteristics, or concepts that are often used to describe or classify people, things, or ideas, such as demographics, attributes, or types.	true
1,269	gen_b20	20	The word "major" is consistently associated with top-level professional sports leagues and competitions.	true
1,093	gen_unconstrained	null	The preposition "at" indicating a location or venue where an event, performance, or action takes place.	true
1,496	gen_b5	5	Prefix indicating negative or intense.	true
1,496	gen_b2	2	Prefix "imp"	true
1,564	gen_b5	5	Hyphenated numbers and quantities.	true
1,037	gen_b40	40	Nouns and adjectives representing various concepts, including identity, orientation, and characteristics, as well as objects, actions, and ideas, often in a context of diversity, inclusivity, and social issues, with a focus on specific categories and attributes.	true
1,496	gen_b10	10	Prefix "imp" indicating negation or intensification of a word.	true
1,564	gen_b2	2	Quantitative values	true
1,564	gen_b10	10	Hyphenated numbers indicating quantities or measurements in text.	true
1,093	gen_b40	40	Preposition "at" indicating a location or venue where an event, performance, or action takes place, often in the context of music, theater, or public speaking, and sometimes in relation to a specific time or schedule.	true
1,496	gen_b20	20	Prefix "imp" often indicating negation, obstruction, or intensity in various words across different contexts and meanings.	true
1,564	gen_b20	20	Hyphenated numbers indicating quantities, measurements, or capacities in various contexts.	true
1,717	gen_b5	5	Comparative adverb indicating greater degree.	true
1,717	gen_b2	2	Comparative intensifier	true
1,564	gen_unconstrained	null	A hyphen-like token separating a number from a unit of measurement, often indicating a quantity or size.	true
1,717	gen_b20	20	Comparative phrases using "more" to convey a greater degree or extent of something.	true
1,269	gen_unconstrained	null	The word "major" is consistently used to describe a top-level league or organization in various sports, such as baseball, soccer, and hockey, often in a formal or official context.	true
1,717	gen_b10	10	Comparative adverb "more" indicating a greater degree or extent.	true
1,269	gen_b40	40	The term "major" is consistently associated with high-level professional sports leagues, such as Major League Baseball (MLB) and Major League Soccer, indicating a top-tier level of competition, often in contrast to minor or junior leagues.	true
1,830	gen_b10	10	The word "dash" often refers to a dashboard or display.	true
1,496	gen_unconstrained	null	Prefix "imp" indicating negation, obstruction, or intensity, often used in words such as "impenetrable", "impasse", "imprecision", "impregnable", "impeding", and others.	true
1,830	gen_b2	2	Dash indicator	true
1,496	gen_b40	40	The prefix "imp" is consistently associated with words that convey a sense of obstruction, hindrance, or negation, such as "impregnable", "impeding", "impossible", and "impostors", indicating a pattern of prefixation with a specific semantic meaning.	true
1,564	gen_b40	40	A hyphen-like token separating a large number from a unit of measurement, often describing the size or capacity of a physical space, object, or quantity, such as square feet, seats, volts, pages, or years.	true
1,830	gen_b5	5	Token "dash" in various contexts.	true
1,717	gen_unconstrained	null	Comparative adverb "more" used to intensify a statement, often indicating a greater degree or extent of something.	true
1,830	gen_b20	20	The token "dash" often represents a dashboard or a sudden movement, sometimes related to technology or vehicles.	true
1,862	gen_b5	5	Emphasis on entirety or completeness.	true
1,862	gen_b2	2	Total inclusion	true
1,717	gen_b40	40	The word "more" is used to intensify comparisons, often in phrases such as "much more" or "far more", to convey a greater degree or extent of something, frequently in formal or written contexts.	true
1,916	gen_b5	5	Abbreviation for news agency.	true
1,916	gen_b2	2	News Agency	true
1,862	gen_b20	20	The word "all" is used to emphasize totality or completeness in various contexts.	true
1,862	gen_b10	10	Emphasis on entirety or completeness with "all" or "most".	true
1,916	gen_b10	10	Abbreviation for news agency or organization names.	true
1,916	gen_b20	20	Abbreviation "IA" often represents a news agency or organization in various contexts.	true
1,830	gen_unconstrained	null	The token "dash" often represents a part of a compound word, typically referring to a control panel or display, such as a dashboard, or indicating a sudden movement or action.	true
2,136	gen_b2	2	Geographic names	true
1,830	gen_b40	40	The token "dash" often represents a noun referring to a dashboard, a panel of instruments or controls, or a sudden movement, and is sometimes used as a prefix in compound words, often related to technology or vehicles.	true
1,862	gen_unconstrained	null	The word "all" is used to emphasize the totality or entirety of something, often in the context of taking action, making decisions, or possessing something.	true
2,155	gen_b2	2	Economic incentive	true
2,136	gen_b5	5	Token "ond" in proper nouns.	true
1,862	gen_b40	40	The word "all" is consistently emphasized, often indicating a comprehensive or exhaustive scope, and is frequently used in phrases such as "take all measures" or "all of the", conveying a sense of totality or completeness.	true
2,136	gen_b10	10	The suffix "ond" in various geographical and proper nouns.	true
1,916	gen_unconstrained	null	Abbreviation for a news agency or organization, often "RIA" ( Rossiyskaya Informatsionnaya Agentura, or Russian Information Agency) or similar, indicating the source of a report or information.	true
2,155	gen_b5	5	Concept of motivation or reward.	true
1,916	gen_b40	40	The token "IA" appears to be an abbreviation for various organizations, news agencies, and institutions, often used in formal or official contexts, such as news reports, government agencies, and international organizations, indicating a reference to a specific entity.	true
2,155	gen_b10	10	The word "incentive" is used to describe motivation or encouragement.	true
2,136	gen_b20	20	The suffix "-ond" appears in various proper nouns, geographical names, and words with specific meanings.	true
2,136	gen_unconstrained	null	The suffix "-ond" appears to be a common thread, often found in proper nouns, geographical locations, and surnames.	true
2,196	gen_b2	2	Government agencies	true
2,155	gen_b20	20	The word "incentive" is consistently associated with motivation or encouragement to take a specific action or make a decision.	true

End of preview. Expand in Data Studio

Length-penalized detection scores for GPT-2 small SAE feature explanations

Data from the project accurate-and-simple-mechanistic-interpretability-explanations (author: Libor "Burny" Burian). It contains 2002 natural-language explanations of 200 SAE latents of GPT-2 small (gpt2-small-res-jb, layer 6), each scored with EleutherAI delphi's detection scorer (Llama-3.1-70B-Instruct), with every per-context prediction.

The project's new score multiplies the above-chance detection score by a length factor: S = max(J, 0) · 2^(−L/H), where J = 2·BA − 1 and L is in o200k_base tokens. S is not stored; it is recomputed from the predictions and the explanation text.

Files

Path	Content
`results/main/scores.jsonl`	One row per (latent, condition): `explanation`, `scorer_model`, `shuffle_seed`, and `samples`, the 50 test contexts with `id`, `activating`, `prediction` and `quantile`.
`results/main/scores_retest.jsonl`, `scores_altscorer.jsonl`	Re-scoring of 50 latents: shuffle seed 43, and Llama-3.3-70B as the scorer.
`results/main/generated_explanations.jsonl`	Llama-3.1-70B explanations at word budgets ~2, 5, 10, 20, 40, and unconstrained.
`results/main/numbers.json`	All analysis numbers (pre-registered hypothesis tests).
`data/processed/`	The Neuronpedia rows of the sampled latents, the sample lists, and the verified negative contexts.
`logs/`	Every LLM prompt and response, untruncated.

Conditions

gpt35 and gpt4omini: Neuronpedia's explanations.
top_token: a baseline of the form the token "X".
random: another latent's explanation (control).
gen_b{2,5,10,20,40} and gen_unconstrained: Llama-3.1-70B explanations from delphi's explainer prompt, with a word budget.
human*: Neuronpedia human explanations.

Main results

Plain detection (J) prefers longer explanations: J goes from 0.464 at ~2 words to 0.577 at ~40 words.
At H = 20, mean S is highest for the shortest explanations, so the pre-registered interior optimum was falsified.
Choosing explanations by S instead of J makes them 39% shorter, for 0.031 held-out J.

See the repository README and paper for the full write-up.

Sources and licenses

This dataset mixes material under its sources' own terms; no single license is asserted.

Explanations and activations come from Neuronpedia's public exports.
Negatives are windows of OpenWebText (CC0).
The SAE is by Joseph Bloom.
Scorer prompts are from delphi (Apache-2.0).

Downloads last month: -

Collection including BurnyCoder/length-penalized-autointerp-gpt2-small

Accurate and simple interpretability explanations

Collection

Length-penalized detection scoring of SAE feature explanations • 1 item • Updated about 3 hours ago