midah/odl-training-data

Training Data Scale Registry

A systematic registry of AI model training data size estimates with evidence profiles.

Dataset Description

This dataset contains structured records of AI models with:

• Token count estimates (min/max/mid)
• Evidence types (E1-E5) and strength (S-High/Medium/Low)
• Uncertainty sources (U1-U5)
• Model metadata (parameters, FLOPs, architecture)
• Raw evidence snippets

Data Collection

Data is collected from:

• Epoch AI datasets
• Hugging Face model cards
• Technical reports and system cards
• Third-party analyses

Inference Methods

Token estimates are derived using:

• Chinchilla scaling law
• Hardware back-calculation
• Parameter ratio heuristics
• Textual token clues
• Third-party analyses

Evidence Profiles

Each model includes an evidence profile indicating:

• Evidence Types: How the estimate was derived
• Evidence Strength: Confidence in the estimate
• Uncertainty Sources: What information is missing

Usage

from datasets import load_dataset

dataset = load_dataset("midah/odl-training-data")

Citation

If you use this dataset, please cite:

Training Data Scale Registry
ODL Research
2025