legal-router

LawAI Baseline Classification Dataset

This repository hosts the dataset used to fine-tune a text classification model under the LawAI project. The baseline run achieved highly optimized classification scores using a modern encoder architecture.

Dataset Summary

The dataset consists of structured legal text samples mapped to classification labels across multiple categories.

• Total Training Examples: 78,358
• Total Validation Examples: 19,590
• Total Schema Columns: ['text', 'label']
• Target Classes: Multi-class classification (including up to index label 18 or greater).

Schema

Baseline Model Performance

The dataset was validated by fine-tuning a microsoft/deberta-v3-base topology (DebertaV2ForSequenceClassification) using a linear learning rate warmup and decay over 6 epochs.

Convergence Metrics

The model reached optimal generalization convergence within the first epoch, remaining highly stable with no signs of overfitting throughout the training cycle:

• Validation Macro F1: 0.99255
• Validation Micro F1: 0.99208
• Final Evaluation Loss: 0.0116
• Training Throughput: ~232 samples/sec on a single GPU environment.

Usage & Loading

You can load this dataset directly via the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/YOUR_DATASET_NAME")

# Inspecting a split
print(dataset['train'][0])