This repository contains the dataset for training and evaluating PureGPT, a lightweight research-focused language model created by Dustin Loring in September 2025. The dataset is designed to align with the OpenAI Harmony chat template and uses the meta-llama/Llama-3.2-1B tokenizer for tokenization.
train.jsonl – 12 examples used for pretraining and fine-tuning.
validation.jsonl – 5 examples for validation (model evaluation during training).… See the full description on the dataset page:
https://huggingface.co/datasets/AIGym/PureGPT_Base.
