Skip to content

Repository files navigation

SoloLLM

Overview

SoloLLM started out as a challenge to see if I could train a base model from scratch on consumer hardware that could reach GPT-2-level performance. After a few model iterations, it had already reached GPT-2 small level performance. At that point the project shifted to a more open small model research project.

The refined goal for this project is to use SoloLLM as a research platform for understanding how to build stronger small language models from scratch with hardware constraints

Model Params Tokens Status Hugging Face
SoloLLM-1.0-152M-Base 151.9M 9.8B Strongest completed SoloLLM 1.0 checkpoint model
SoloLLM-1.0-123M-Base 123.6M 9.8B Smaller-than-GPT-2 comparison checkpoint model

Current Findings

  • SoloLLM-1.0-152M-Base is the strongest checkpoint so far. It beats GPT-2 small across the fixed base-model evaluation suite, but it is slightly larger than GPT-2 small.
  • SoloLLM-1.0-123M-Base is slightly smaller than GPT-2 small. It wins the normalized multiple-choice suite and several external checks, but it does not beat GPT-2 small across every metric.
  • Data mix seemed to have the largest impact on model performance. A reasoning-heavy mix improved some multiple-choice and repetition metrics, but hurt broad language-modeling perplexity. A broader blend with dropout 0.0 produced the best overall result.

Documents


Benjamin Maxwell


MIT License

About

GPT-2-class language models trained from scratch in PyTorch on one RTX 3090, with 10B-token data curation and full GPT-2 comparisons.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Contributors

Languages