06RESEARCH PAPERPreprint
Mercury-Viewer: A Self-Contained Browser-Based Interactive 3D Visualization of LLM Internal Sensor Maps
發表日期
摘要
ABSTRACTZenodo 原文v2 corrigendum (2026-05-21): Bibliography polished to canonical forms — Bycroft (December 2023) year corrected, full author lists added for Bricken et al. (24 authors) and Carter et al. (5 authors), DOIs added for Vig (10.18653/v1/P19-3007) and Activation Atlas (10.23915/distill.00015). Self-references aligned to canonical author name. Author byline corrected from 'norika' to 'Chen, Ho Yiing'. No changes to data, viewer, or claims.
This record releases Mercury-Viewer, a single HTML file (~2 MB) that opens in any browser and shows the inside of a transformer language model as a 3D point cloud. No install, no server, no GPU, no API key.
Mercury-Viewer reads activation data from the Mercury sensor grid and draws it in cylindrical coordinates: azimuth is the hidden dimension index, radius is the activation quantile, height is how many prompts activated a given unit. Every point keeps a full (L, D, Q) coordinate, so hovering reveals the exact location, and the picture is invertible.
Alongside the viewer, this record releases the Mercury-3B-7B observation dataset: 21,694 fired-cell records and 28,654 visible Rosen-bridge co-firing edges, captured by running ten multilingual prompts through qwen2.5:3B and qwen2.5:7B on a single RTX 3060. Code MIT, data CC-BY-4.0.
Companion records: Mercury method paper (10.5281/zenodo.20313154), and forthcoming Mercury discovery and scaling preprints.
全文 PDF
FULL TEXT1 個檔- 下載 PDFpaper4-mercury-viewer.pdf・119 KB・本站
PDF 與 Zenodo 上的檔案相同(大小與 md5 已核對)。
論文資料
DETAILSZenodo 紀錄 20313150- 作者
- Chen, Ho Yiing
- 發表日期
- 資源類型
- Preprint
- DOI
- 10.5281/zenodo.20313150
- Zenodo
- zenodo.org/records/20313150
- 授權
- Creative Commons Attribution 4.0 International
- 版本
- 1.1.0
作者 Chen, Ho Yiing 是織田紀香的本名(陳禾穎)英文拼寫。標題、摘要、關鍵字與授權照 Zenodo 原文。
關鍵字
KEYWORDS16 個- LLM interpretability
- mechanistic interpretability
- transformer visualization
- activation analysis
- sensor grid
- neural network observability
- Qwen2.5
- browser-based visualization
- Three.js
- open dataset
- consumer hardware
- independent research
- Mercury
- Rosen bridge
- co-firing
- multilingual NLP