Model Extraction Playground

Interactive analysis of LLM performance on charity document data extraction

Model Leaderboard

F1 Score Overview

Same models, order, and F1 values as the leaderboard (scroll to see all).

Provider Summary

Aggregated stats per provider (active models only, F1 > 0). Fields and cost are averages per model.

Scoring Methodology

Field-Level Accuracy Heatmap

Each cell shows how accurately a model extracted that field across all 11 documents. Click a cell for details.

Best Model per Field

Exact-match accuracy among functional models (F1 > 0).

Field Difficulty Ranking

Average exact-match accuracy across functional models — table and chart share the same rows.

Document × Model Accuracy Heatmap

Each cell shows accuracy for a specific document/model pair.

Document Difficulty

Average exact-match accuracy across functional models (error rows excluded from the average). Table and chart share the same rows.

Error Category Breakdown per Model

Exact-match correct / wrong / missing shares — same order and values as the chart.

Correct
Wrong value
Missing

Stacked share (exact-match)

Same models and percentages as the table (scroll to see all).

Common Error Patterns

Per-Document Field Comparison

Decision Helper

Key Insights

Improvement Suggestions

Provider Comparison

Active models only — same aggregates as Rankings → Provider Summary (Avg/Best F1).

Tier Distribution by Provider

How models are distributed across pricing tiers per provider.

Cost Efficiency (F1 per $)

Higher is better — F1 score achieved per dollar spent (active models with cost data).

Modality Breakdown

Text-only vs multimodal model distribution and performance per provider.

Model Catalog

All configured models with capabilities, pricing, and performance. Filter by provider.

Project Evolution Timeline

How this extraction benchmark evolved — from raw data preparation to a multi-provider, multi-model scored leaderboard.

Numbers at a Glance

Cost & Speed Summary

From extraction_stats.csv — actual observed time and cost per model (where available).