Neat Vision: Inspect NLP attention and predictions

Understand the project

Practical overview

Use it when you have sentence-level self-attention outputs and want to inspect NLP model behavior or produce an attention visualization.

NLP researchers preparing attention visualizations for analysis or academic papers.
Developers inspecting predictions and debugging sentence-level NLP models.

Visualize attention scores alongside the tokens they correspond to.
Inspect classification posteriors and regression errors alongside model predictions.
Export attention visualizations as SVG.

Before you choose

Input data must include tokenized text and attention scores; prediction fields vary by task.
Classification posterior values are optional and may be supplied as unnormalized logits.
SVG export may require conversion to PDF for a paper workflow.
The repository is public and identifies an MIT license; Vue is the listed primary language.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Format model results as the input data Neat Vision uses to visualize attention and predictions.

A JSON data file containing tokenized text, attention scores and the task's required labels and predictions.
For classification tasks, an optional label file can map class labels to descriptions.

Choose regression, multi-class classification or multi-label classification, the documented sentence-level tasks.
Create a data record with tokenized text, attention weights and a unique sample ID.
For regression, include the actual label and predicted value; for classification, include the label and prediction. Add posterior values for classification if available.
Use the documented input format for the chosen task; posterior values may be unnormalized logits, which Neat Vision normalizes.

A task-formatted JSON data file containing the inputs Neat Vision needs to visualize attention and, when included, model predictions.

The documented scope covers sentence-level self-attention, not document-level hierarchical or sequence-to-sequence models.
Attention weights are required; prediction-related posterior values are optional for classification.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 17 October 2026
Original website title: cbaziotis/neat-vision · Official homepage title · retained in its source language