Attention Atlas

Back

An interactive visualization of Transformer internals with a focus on attention mechanisms.

Mode

Basic Advanced

Compare Modes

Models

Prompts

Model Family

Model Configuration

Model Family - B

Model Config - B

Input Text

Browse...

Batch ModeUpload a CSV or JSON to analyse multiple sentences at once.

Drop CSV or JSON file here, or click to browse

Browse...

Processing...

Attention Atlas

Back

Detect and analyze social bias in text using GUS-Net neural NER.

Compare Modes

Models

Prompts

Bias Detection Model

Detect Model - B

Sensitivity Threshold

UNFAIR0.50

GEN0.50

STEREO0.50

UNFAIR0.50

GEN0.50

STEREO0.50

Input Text

Browse...

Batch ModeUpload a CSV or JSON to analyse multiple sentences at once

Drop CSV or JSON file here, or click to browse

Browse...

Metric Explanation

Inter-Sentence Attention Details

Sentence X (Target)

Sentence Y (Source)

ISA Score:

What does this represent?

This value is the average fraction of Sentence X's attention that lands on Sentence Y, with attention averaged over all layers and heads and attention to special tokens excluded. It ranges from 0 to 1 and reads as a share of X's attention budget.

Interpretation

High ISA → strong dependency across sentences (semantic or syntactic connection)
Low ISA → weak or no cross-sentence influence

Token-to-Token Attention

Auditor Notebook

Record one analytical move per entry. Stored in this session and persisted to disk; export when you finish.

New entry

Title is optional. The five thesis elements are saved: hypothesis, automatic conditions, automatic signals, uncertainty, and next steps.

Audit case ID (optional)

Group related entries under one investigation, e.g. "crows-pairs-race-bert-2026-05".

Title (optional)

A short label so the entry is easy to find later.

Hypothesis

What you expect the model to do, and why.

Uncertainty acknowledged

What this evidence cannot decide, and what could overturn it.

Next steps

Concrete follow-ups: another model, another axis, another prompt, a control.

Conditions and signals captured automatically

Entries

Persisted to downloads/sessions/auditor_notebook.json.

Export Markdown Export JSON Export CSV