{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "data": { "@id": "cr:data", "@type": "@json" }, "dataBiases": "rai:dataBiases", "dataCollection": "cr:dataCollection", "dataCollectionType": "cr:dataCollectionType", "dataCollectionMissingData": "cr:dataCollectionMissingData", "dataCollectionRawData": "cr:dataCollectionRawData", "dataCollectionTimeFrameStart": "cr:dataCollectionTimeFrameStart", "dataCollectionTimeFrameEnd": "cr:dataCollectionTimeFrameEnd", "dataPreprocessingProtocol": "cr:dataPreprocessingProtocol", "dataType": { "@id": "cr:dataType", "@type": "@vocab" }, "dct": "http://purl.org/dc/terms/", "extract": "cr:extract", "field": "cr:field", "fileObject": "cr:fileObject", "fileProperty": "cr:fileProperty", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isEnumeration": "cr:isEnumeration", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "personalSensitiveInformation": "rai:personalSensitiveInformation", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform", "annotationsPerItem": "cr:annotationsPerItem", "annotatorDemographics": "cr:annotatorDemographics", "machineAnnotationTools": "cr:machineAnnotationTools", "rai": "http://mlcommons.org/croissant/rai/1.0/", "dataLimitations": "rai:dataLimitations", "dataUseCases": "rai:dataUseCases", "dataSocialImpact": "rai:dataSocialImpact", "hasSyntheticData": "rai:hasSyntheticData", "prov": "http://www.w3.org/ns/prov#", "wasDerivedFrom": "prov:wasDerivedFrom", "wasGeneratedBy": "prov:wasGeneratedBy" }, "@type": "sc:Dataset", "name": "PoolBench", "description": "PoolBench is a benchmark for evaluating hidden-state pooling strategies for activation steering across 17 text concepts for language models. Each concept corpus contains 1,000 passages per class (700 train / 300 test) sourced from real public datasets on HuggingFace. No passages are LLM-generated. The benchmark measures three dimensions: (D1) linear separability (AUROC) of pooled concept vectors, (D2) steered concept prevalence (SCP) at injection, and (D3) output-level disentanglement from neighbouring concepts.", "url": "https://huggingface.co/datasets/nips234678/poolbench", "conformsTo": "http://mlcommons.org/croissant/1.1", "citeAs": "@misc{poolbench2026, title={PoolBench: Evaluating Pooling Strategies for Activation Steering Vectors}, author={Anonymous}, year={2026}}", "license": "https://creativecommons.org/licenses/by/4.0/", "inLanguage": "en", "keywords": [ "activation steering", "language model interpretability", "pooling strategies", "hidden states", "benchmark", "NLP", "steering vectors" ], "creator": { "@type": "Person", "name": "Anonymous" }, "datePublished": "2026-05-04", "version": "1.0.0", "distribution": [ { "@type": "http://mlcommons.org/croissant/FileObject", "@id": "poolbench-repo", "name": "PoolBench HuggingFace repository", "contentUrl": "https://huggingface.co/datasets/nips234678/poolbench", "encodingFormat": "application/x-tar", "sha256": "6012b31f4d01582dc95943c9fc91b6a05d638ecac8143f4ac47a788b80d7dda3" }, { "@type": "http://mlcommons.org/croissant/FileSet", "@id": "poolbench-corpus-jsonl", "name": "PoolBench JSONL corpus files", "description": "Per-concept JSONL files for all 17 concepts, each with train_pos, train_neg, test_pos, and test_neg splits. Each JSON line is one passage record.", "containedIn": { "@id": "poolbench-repo" }, "includes": "data/corpora/**/*.jsonl", "encodingFormat": "application/jsonlines" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "passage", "name": "Passage record", "description": "A single text passage from the PoolBench corpus with its concept label, metadata, and provenance.", "field": [ { "@type": "cr:Field", "@id": "passage/text", "name": "text", "description": "The passage text, 300–500 tokens (200–500 for depression) as tokenised by meta-llama/Meta-Llama-3.1-8B.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.text" } } }, { "@type": "cr:Field", "@id": "passage/label", "name": "label", "description": "Binary concept label. 1 = concept present (positive class); 0 = concept absent (negative class).", "dataType": "sc:Integer", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.label" } } }, { "@type": "cr:Field", "@id": "passage/concept", "name": "concept", "description": "Which of the 17 benchmark concepts this passage belongs to.", "dataType": "sc:Text", "isEnumeration": true, "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.concept" } } }, { "@type": "cr:Field", "@id": "passage/domain", "name": "domain", "description": "Source domain tag (e.g. academic, news, social, fiction, review). At least 3 distinct domains per concept are required by the benchmark corpus rules.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.domain" } } }, { "@type": "cr:Field", "@id": "passage/token_count", "name": "token_count", "description": "Token count of the passage as returned by the meta-llama/Meta-Llama-3.1-8B tokenizer.", "dataType": "sc:Integer", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.token_count" } } }, { "@type": "cr:Field", "@id": "passage/source_dataset", "name": "source_dataset", "description": "Original HuggingFace dataset ID from which the passage was sourced.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.source_dataset" } } }, { "@type": "cr:Field", "@id": "passage/matched_pair_id", "name": "matched_pair_id", "description": "Pair identifier for matched-pair concepts (hedging, legal_formality, causation, contrast, conditionality, negation_density). Null for independently-sampled concepts (frustration, imdb_sentiment, toxicity, depression, academic_tone, code_docs, bureaucratic, narrative, deference, planning, numerical_precision).", "dataType": "sc:Text", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.matched_pair_id" } } }, { "@type": "cr:Field", "@id": "passage/split", "name": "split", "description": "Dataset split: train (700/class) or test (300/class). Split is fixed before any model is run and never changed.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "poolbench-corpus-jsonl" }, "extract": { "jsonPath": "$.split" } } } ] } ], "cr:dataCollection": "All passages are sourced from existing public datasets on HuggingFace. No passage is generated by a language model. Source datasets include: ArXiv abstracts (gfissore/arxiv-abstracts-2021), BillSum (FiscalNote/billsum), CC-News (cc_news), LexGLUE/SCOTUS and EurLex (lex_glue), Yelp reviews (Yelp/yelp_review_full), IMDb reviews (yin001/imdb_dataset_positive_negative), NuminaMath-CoT (AI-MO/NuminaMath-CoT), MetaMathQA (meta-math/MetaMathQA), PubMedQA (qiaojin/PubMedQA), CodeSearchNet (code_search_net), Reddit posts (sentence-transformers/reddit), Jigsaw Civil Comments (google/civil_comments), Davidson hate-speech corpus (tdavidson/hate_speech_offensive), Surge-AI toxicity CSV (github.com/surge-ai/toxicity), depression Reddit data (mrjunos/depression-reddit-cleaned, dlb/mentalreddit), Intel Polite-Guard (Intel/polite-guard), WikiHow (gursi26/wikihow-cleaned), WritingPrompts (euclaise/writingprompts), Wikipedia (wikimedia/wikipedia).", "cr:dataCollectionType": "Corpus construction from existing public datasets. Passages are filtered by concept-specific lexical, syntactic, or semantic criteria and validated against token-length and label-purity rules.", "cr:dataCollectionRawData": "Raw source data was obtained from HuggingFace Hub dataset repositories linked above. No original data collection from the web was performed.", "cr:dataCollectionTimeFrameStart": "2026-01-01", "cr:dataCollectionTimeFrameEnd": "2026-04-30", "cr:dataCollectionMissingData": "Passages failing the 300–500 token filter, matched-pair token-diff > 25, or seed-word contamination checks are discarded. For concepts where fewer than 1,000 clean passages per class were findable, the corpus is smaller than target and documented in the corpus README.", "cr:dataPreprocessingProtocol": "1. Passages chunked to 300–500 tokens (meta-llama/Meta-Llama-3.1-8B tokenizer). 2. Concept-specific filters applied (seed-word, regex, spaCy dependency parse). 3. MD5 deduplication within and across splits. 4. Label contamination check: negative passages must not contain positive-class seed words (word-boundary matching). 5. Train/test split stratified by label and fixed before any model experiment. 6. Matched-pair token-diff ≤ 25 enforced for 6 syntactic/sparse-lexical concepts. All filter functions are published in the repository for exact reproducibility.", "rai:dataBiases": "1. Source dataset demographic skew: Reddit-sourced corpora (depression, frustration, planning, narrative) reflect English-language Reddit demographics, which skew young, male, and North American. 2. Civil Comments / Davidson: historical annotation biases in hate-speech corpora are documented in those datasets' original papers. 3. IMDb sentiment corpus is single-domain (movie reviews) and may not generalise to other positive-sentiment domains. 4. Deference corpus uses 2 domains (Intel Polite-Guard for customer service; Reddit academic text as social_academic fallback). Advisory: ≥3 domains recommended for a future release. 5. All passages are English-only. Non-English adaptations of PoolBench are out of scope for this release.", "cr:annotationsPerItem": "Each passage carries one binary concept label (0 or 1). Labels are assigned by: (a) the passage's source dataset's original labels (for datasets with existing concept-level annotation: Davidson hate-speech corpus, Intel Polite-Guard), or (b) deterministic filter functions based on lexical/syntactic criteria (all other concepts). No crowd-sourced re-annotation was performed on PoolBench passages.", "cr:annotatorDemographics": "No human annotators were used in PoolBench corpus construction. All labels derive from: (1) existing expert-annotated source datasets (see dataCollection above), or (2) deterministic rule-based filters implemented by the paper authors. No AMT, Prolific, or similar crowd-sourcing was used.", "cr:machineAnnotationTools": "spaCy en_core_web_sm: syntactic dependency parse for negation_density (token.dep_=='neg'), conditionality, causation concepts. meta-llama/Meta-Llama-3.1-8B tokenizer: token-length computation for all passages. Python regex: numerical_precision filter, text cleanup, seed-word detection.", "rai:personalSensitiveInformation": "PoolBench passages are sourced from public internet text (Reddit, Yelp reviews, news articles, Wikipedia, academic publications). Passages are not linked to identifiable individuals. Usernames and identifying metadata are not included in the corpus. Two concepts contain sensitive content: (1) depression — passages express mental-health distress; these are sourced from publicly posted Reddit text where users chose to share this content publicly. (2) toxicity — passages contain hate speech, offensive language, or hostile content; these are sourced from existing research corpora with established ethical review. Steering vectors and steered outputs for both sensitive concepts are not released publicly (see ethics section of the paper).", "isAccessibleForFree": true, "measurementTechnique": "Logistic regression probe AUROC (D1), steered concept prevalence via frozen BERT classifier (D2), output-level disentanglement ratio (D3).", "rai:dataLimitations": "1. All passages are English-only; PoolBench should not be used to evaluate pooling strategies for non-English or multilingual steering vectors without re-construction in the target language. 2. Depression and toxicity corpora reflect Reddit and Twitter demographics (English-speaking, skewed young/North American/male); steering vectors trained on these corpora may not generalise to clinical populations or other cultural contexts. 3. IMDb sentiment is single-domain (movie reviews); concept validity is established for review-level polarity but not for cross-domain sentiment generalisation. 4. Deference corpus uses 2 domains; wider generalisation across formal writing registers has not been validated. 5. The benchmark fixes construction method at DiffMean for the primary leaderboard; conclusions about pooling strategy ranking may not hold under substantially different construction methods (see the paper appendix for a Spearman rank stability check). 6. Passages are 300–500 tokens; findings may not extrapolate to very short (< 100 token) or very long (> 2,000 token) context regimes.", "rai:dataUseCases": "Established valid uses: (1) Benchmarking and ranking activation steering vector pooling strategies by linear separability (AUROC on logistic regression probe). (2) Analysing the relationship between concept type (sparse-lexical, syntactic, register, semantic-abstract, dense-lexical) and pooling strategy performance. (3) Evaluating steered concept prevalence in LLM outputs across 3 generative models. (4) Measuring output-level concept disentanglement. Not validated for: (1) Fine-tuning language models (corpus is not a general-purpose text dataset). (2) Hate speech or toxicity detection as a standalone classifier benchmark (toxicity corpus is drawn from research corpora already used for that purpose). (3) Clinical mental health assessment (depression corpus is Reddit text, not clinical notes). (4) Cross-lingual or multilingual activation steering.", "rai:dataSocialImpact": "Positive impacts: PoolBench enables systematic comparison of activation steering methods, advancing the interpretability toolchain for large language models. Verified pooling strategies could make steering more reliable for safety-relevant applications such as reducing harmful outputs. Negative / risk: (1) Improved steering techniques could also be misused to more reliably inject undesirable concepts into LLM outputs. (2) The toxicity concept corpus contains real hate speech and offensive language; releasing the corpus publicly could provide adversarial content. Mitigations: steering vectors and steered outputs for the depression and toxicity concepts are not released publicly in the primary artefact release. The corpus is released under CC-BY 4.0 with an acceptable-use note requesting that downstream users not use these sub-corpora to train classifiers intended for harassment or surveillance. (3) Depression corpus contains Reddit posts expressing mental-health distress; these are publicly posted text but may be sensitive if re-identified.", "prov:wasDerivedFrom": [ { "@type": "sc:Dataset", "name": "ArXiv Abstracts 2021", "url": "https://huggingface.co/datasets/gfissore/arxiv-abstracts-2021", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "BillSum", "url": "https://huggingface.co/datasets/FiscalNote/billsum", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "CC-News", "url": "https://huggingface.co/datasets/cc_news", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "LexGLUE (SCOTUS + EurLex)", "url": "https://huggingface.co/datasets/lex_glue", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "Yelp Review Full", "url": "https://huggingface.co/datasets/Yelp/yelp_review_full", "license": "https://huggingface.co/datasets/Yelp/yelp_review_full" }, { "@type": "sc:Dataset", "name": "IMDb Dataset (positive/negative)", "url": "https://huggingface.co/datasets/yin001/imdb_dataset_positive_negative", "license": "https://huggingface.co/datasets/yin001/imdb_dataset_positive_negative" }, { "@type": "sc:Dataset", "name": "NuminaMath-CoT", "url": "https://huggingface.co/datasets/AI-MO/NuminaMath-CoT", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "MetaMathQA", "url": "https://huggingface.co/datasets/meta-math/MetaMathQA", "license": "https://opensource.org/licenses/MIT" }, { "@type": "sc:Dataset", "name": "PubMedQA", "url": "https://huggingface.co/datasets/qiaojin/PubMedQA", "license": "https://opensource.org/licenses/MIT" }, { "@type": "sc:Dataset", "name": "CodeSearchNet", "url": "https://huggingface.co/datasets/code_search_net", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "Reddit (sentence-transformers)", "url": "https://huggingface.co/datasets/sentence-transformers/reddit", "license": "https://huggingface.co/datasets/sentence-transformers/reddit" }, { "@type": "sc:Dataset", "name": "Jigsaw Civil Comments", "url": "https://huggingface.co/datasets/google/civil_comments", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "Davidson Hate Speech and Offensive Language", "url": "https://huggingface.co/datasets/tdavidson/hate_speech_offensive", "license": "https://opensource.org/licenses/MIT" }, { "@type": "sc:Dataset", "name": "Surge-AI Toxicity", "url": "https://github.com/surge-ai/toxicity", "license": "https://www.apache.org/licenses/LICENSE-2.0" }, { "@type": "sc:Dataset", "name": "Depression Reddit Cleaned", "url": "https://huggingface.co/datasets/mrjunos/depression-reddit-cleaned", "license": "https://huggingface.co/datasets/mrjunos/depression-reddit-cleaned" }, { "@type": "sc:Dataset", "name": "MentalReddit", "url": "https://huggingface.co/datasets/dlb/mentalreddit", "license": "https://creativecommons.org/publicdomain/zero/1.0/" }, { "@type": "sc:Dataset", "name": "Intel Polite-Guard", "url": "https://huggingface.co/datasets/Intel/polite-guard", "license": "https://www.apache.org/licenses/LICENSE-2.0" }, { "@type": "sc:Dataset", "name": "WikiHow", "url": "https://huggingface.co/datasets/gursi26/wikihow-cleaned", "license": "https://creativecommons.org/licenses/by-nc-sa/4.0/" }, { "@type": "sc:Dataset", "name": "WritingPrompts", "url": "https://huggingface.co/datasets/euclaise/writingprompts", "license": "https://creativecommons.org/licenses/by/4.0/" }, { "@type": "sc:Dataset", "name": "Wikipedia", "url": "https://huggingface.co/datasets/wikimedia/wikipedia", "license": "https://creativecommons.org/licenses/by-sa/4.0/" } ], "prov:wasGeneratedBy": [ { "@type": "prov:Activity", "name": "Data Collection", "cr:activityType": "Collection", "description": "Source passages were streamed from 20 existing public HuggingFace datasets (listed in prov:wasDerivedFrom). No original web crawl was performed. For each of the 17 benchmark concepts a concept-specific filter was applied: lexical seed-word regex, syntactic dependency parse (spaCy en_core_web_sm), or the source dataset's original label. All 17 concepts required at least 3 distinct source domains; passages were drawn from news, academic, social, fiction, review, legal, and code domains as appropriate.", "prov:startedAtTime": "2026-01-01", "prov:endedAtTime": "2026-04-30", "prov:wasAssociatedWith": [ { "@type": "prov:Agent", "name": "Paper authors (anonymous)", "cr:agentType": "HumanAgent" }, { "@type": "prov:Agent", "name": "spaCy en_core_web_sm (syntactic filter)", "cr:agentType": "SoftwareAgent" } ] }, { "@type": "prov:Activity", "name": "Preprocessing and Filtering", "cr:activityType": "Preprocessing", "description": "Each candidate passage was: (1) chunked to 300–500 tokens using the meta-llama/Meta-Llama-3.1-8B tokenizer (200–500 for the depression concept); (2) checked for seed-word contamination in negative-class passages (word-boundary regex matching); (3) MD5-deduplicated within split, across splits same-label, and across splits cross-label; (4) for 6 syntactic/sparse-lexical concepts (hedging, legal_formality, causation, contrast, conditionality, negation_density), positive and negative passages were rewritten into matched pairs using rule-based marker removal, with a token-diff ≤ 25 constraint enforced. Passages failing any filter were discarded.", "prov:startedAtTime": "2026-01-01", "prov:endedAtTime": "2026-04-30", "prov:wasAssociatedWith": [ { "@type": "prov:Agent", "name": "Paper authors (anonymous)", "cr:agentType": "HumanAgent" }, { "@type": "prov:Agent", "name": "meta-llama/Meta-Llama-3.1-8B tokenizer", "cr:agentType": "SoftwareAgent" }, { "@type": "prov:Agent", "name": "Python regex and rule-based pipeline (open-source, published in repository)", "cr:agentType": "SoftwareAgent" } ] }, { "@type": "prov:Activity", "name": "Annotation and Split Assignment", "cr:activityType": "Annotation", "description": "Each passage received one binary concept label (0 or 1) via one of two routes: (a) the source dataset's existing expert annotation (Davidson hate-speech corpus; Intel Polite-Guard) was carried over directly; (b) for all other concepts, labels were assigned deterministically by the published filter functions — no human re-annotation or crowd-sourcing was performed. A 70/30 stratified train/test split was then fixed by label before any model experiment was run and was never subsequently changed. No AMT, Prolific, or similar crowd-sourcing platform was used. No LLM was used to generate or label any passage.", "prov:startedAtTime": "2026-01-01", "prov:endedAtTime": "2026-04-30", "prov:wasAssociatedWith": [ { "@type": "prov:Agent", "name": "Paper authors (anonymous)", "cr:agentType": "HumanAgent" } ] } ], "variableMeasured": [ "Concept label (binary)", "Token count", "Source domain", "Matched pair identifier" ], "rai:hasSyntheticData": false }