Axiomic Labs

Axiomic Labs is an independent research lab that trains small language models from scratch. We believe that small language models are an underserved area of research in the AI industry, and that further research into this field could bring great benefits to mankind, by improving LLM guardrails and monitoring, model organisms, and true-edge inference.

Our first model outperformed OpenAI's GPT-3 at HellaSwag using 20x less training data. Six months later we introduced GPT-X3-150M, which is the best model in the world under 150M parameters*.

What we work on

• The GPT-X family of open-weight language models
• LogicMark and ArithMark, reasoning benchmarks for models too small for MMLU
• Training datasets, including SFTset-SLM and FactBase-DCLM
• The Open SLM Leaderboard, which scores small models from every lab with the same method

Latest release

Our third-generation model, GPT-X3-150M, ranks first on the Open SLM Leaderboard. It is the first model under 150M parameters to score above 30 on the leaderboard's Intelligence Index.

* According to a weighted average of chance-normalised scores on HellaSwag, ARC (Easy and Challenge combined), PIQA and ArithMark-3, weighted 1, 1, 1 and 0.65 respectively. This is the Intelligence Index used by the Open SLM Leaderboard.