CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

This is a Hugging Face Space that visualizes ICONCLASS predictions from the davanstrien/iconclass-vlm model compared to ground truth labels. It's a static web application that fetches and displays data from the davanstrien/iconclass-sft-predictions dataset.

The model being evaluated is a fine-tuned vision-language model (based on Qwen/Qwen2.5-VL-3B-Instruct) that automatically classifies art and cultural heritage images using Iconclass notation. The visualization shows how well the model's predictions match the ground truth labels.

Architecture

The project consists of a single-page web application:

index.html: Main application with embedded CSS and JavaScript
style.css: Additional styles (currently minimal, most styles are inline in index.html)
Uses the Hugging Face Datasets Server API to fetch data

Key Implementation Details

Data Source

Dataset: davanstrien/iconclass-sft-predictions
Config: default
Split: test
API endpoint: https://datasets-server.huggingface.co/rows

Data Structure

Each row contains:

images: Array of image objects with src URLs
iconclass-prediction: Raw prediction text
iconclass-predictions-parsed: Parsed prediction labels array
iconclass-gt-parsed: Parsed ground truth labels array

Core Functionality

Lazy loading with pagination (10 images at a time)
Infinite scroll support
Visual comparison between predictions and ground truth
Match detection and scoring
Invalid label detection (labels containing "not a valid" or "invalid")

Development Notes

Since this is a static Hugging Face Space (sdk: static), there's no build process or backend. All changes are made directly to the HTML/CSS files and are immediately reflected when deployed to Hugging Face Spaces.

Important: The README.md file is not displayed in static Spaces. Any documentation or description about the Space must be added directly to the index.html file to be visible to users.

To test locally, simply open index.html in a web browser.