{"paper":{"title":"Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Gemma Scope releases JumpReLU sparse autoencoders for every layer and sub-layer of the Gemma 2 2B and 9B models.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Anca Dragan, Arthur Conmy, J\\'anos Kram\\'ar, Lewis Smith, Neel Nanda, Nicolas Sonnerat, Rohin Shah, Senthooran Rajamanoharan, Tom Lieberum, Vikrant Varma","submitted_at":"2024-08-09T16:06:42Z","abstract_excerpt":"Sparse autoencoders (SAEs) are an unsupervised method for learning a sparse decomposition of a neural network's latent representations into seemingly interpretable features. Despite recent excitement about their potential, research applications outside of industry are limited by the high cost of training a comprehensive suite of SAEs. In this work, we introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models. We primarily train SAEs on the Gemma 2 pre-trained models, but additionally release SAEs"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That standard reconstruction and sparsity metrics are sufficient to establish that the released SAEs will be useful for downstream interpretability and safety research.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Gemma Scope supplies trained sparse autoencoders for all layers of Gemma 2 2B and 9B plus select 27B layers, with public weights and benchmark scores.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Gemma Scope releases JumpReLU sparse autoencoders for every layer and sub-layer of the Gemma 2 2B and 9B models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6d98d96415b23724b95c2689eac0d071078d615623a41c02a648f34ff01e02c2"},"source":{"id":"2408.05147","kind":"arxiv","version":2},"verdict":{"id":"ab479788-8b9f-4331-a394-fd184d075edf","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:43:33.135686Z","strongest_claim":"We introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models.","one_line_summary":"Gemma Scope supplies trained sparse autoencoders for all layers of Gemma 2 2B and 9B plus select 27B layers, with public weights and benchmark scores.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That standard reconstruction and sparsity metrics are sufficient to establish that the released SAEs will be useful for downstream interpretability and safety research.","pith_extraction_headline":"Gemma Scope releases JumpReLU sparse autoencoders for every layer and sub-layer of the Gemma 2 2B and 9B models."},"references":{"count":14,"sample":[{"doi":"","year":2024,"title":"Towards Automated Circuit Discovery for Mechanistic Interpretability","work_id":"8407204b-a28a-4ba8-ae13-e7c04d3497c4","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Gemini: A Family of Highly Capable Multimodal Models","work_id":"83f7c85b-3f11-450f-ac0c-64d9745220b2","ref_index":2,"cited_arxiv_id":"2312.11805","is_internal_anchor":true},{"doi":"","year":2024,"title":"Universal neurons in gpt2 language models","work_id":"eb5860a6-61cb-43e3-8b19-4ddbd580abe0","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","ref_index":4,"cited_arxiv_id":"1412.6980","is_internal_anchor":true},{"doi":"10.18653/v1/2023.blackboxnlp-1","year":2005,"title":"Efficient Estimation of Word Representations in Vector Space","work_id":"59edaa01-a696-45b3-9a08-5eae777a799e","ref_index":5,"cited_arxiv_id":"1301.3781","is_internal_anchor":true}],"resolved_work":14,"snapshot_sha256":"9b45b0ce42befd474e566c7d159c4fc587b6cbdc987f566f554ce66e80424dee","internal_anchors":4},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}