DataValidation

Data Validation Tool for Claude-3 Haiku Fine-Tuning Training and Validation Datasets

This tool provides a simple and efficient way to validate JSONL format data for Claude-3 Haiku Fine-Tuning training and validation datasets. It supports both local files and data stored in Amazon S3.

Contents

1. data_validator.py: The core Python script containing validation logic

2. Claude-3 Haiku Fine-Tuning Training and Validation Data Validator.ipynb: A Jupyter notebook interface for easy use of the validation tool

Features

1. Validates JSONL format for training and validation datasets

2. Supports both local files and S3 stored data

3. Checks for file size limits:

   • Training data: Max 10GB
   • Validation data: Max 1GB

4. Validates line counts:

   • Training data: 32 to 10,000 lines
   • Validation data: 32 to 1,000 lines
   • Total (training + validation): Max 10,000 lines

5. Validates data structure and content for each entry

6. Estimates and checks token counts per entry: Max 32,000 tokens

7. Checks for Anthropic's reserved keywords in prompts:

   • Ensures "\nHuman:" and "\nAssistant:" do not appear in prompts
   • Note: Variations without colons (e.g., "\nHuman" or "\nAssistant") are allowed