Shakespeare SAE Feature — Activation Matrices and Summary Database

This record contains the sparse-autoencoder (SAE) activation data underlying a study of 38 Shakespeare plays. The Folger Shakespeare TEI texts were divided into 4,600 chunks and processed with two language models at three layers each (Gemma-2-2B: layers 4, 11, 22; Gemma-3-4B: layers 9, 17, 29), using Gemma Scope and Gemma Scope 2 SAEs of width 16,384. The 20 files comprise: six firing-rate matrices and six mean-activation matrices (4,600 chunks x 16,384 features each), a derived SQLite summary database, marker-feature records with their strongest-activating tokens, residual-stream samples for ten chunks per model retained for independent recomputation, a snapshot of the Neuronpedia API responses for the 505 features selected for inspection in the paper (fetched 2026-10-12), sanitized run metadata, SHA-256 checksums, and a README describing every file and definition. Read README_zenodo.md first: it defines firing, firing rate, and mean activation, and explains how rows correspond to the chunk table. Code, the chunk table, and aggregate results are maintained in the GitHub repository listed under Related works. The accompanying paper (Jinmonkon 2026) is forthcoming and its citation will be added to this record. 本レコードは、シェイクスピア戯曲38作品を対象とした研究の、SAE 活性データ一式です。Folger Shakespeare の TEI 本文を4,600チャンクに分割し、2つの言語モデルの各3層で特徴量の発火を記録しました。論文の実体確認の対象として選んだ505個の特徴量について、Neuronpedia の API 応答の写し(2026-10-12 取得)も含みます。各ファイルの説明と用語の定義は README_zenodo.md にあります。コードとチャンク表、集計結果は Related works の GitHub リポジトリで管理しています。