Datamule will soon be open sourcing models on Hugging Face. Model training is currently focused on cheap, small models that can spin up on demand in seconds.
Datamule Parser Text 1 2B will be a fine-tuned Gemma 4 model designed to cheaply parse HTML, searchable PDFs, and text files in conjunction with doc2struct, a fast algorithm to parse documents. The model receives the output of the algorithm, then uses tool calls to fix mistakes.
It is designed to run on demand, fitting onto L4 GPUs.
Datamule Parser XBRL 1 *B will be a fine-tuned model to cheaply extract untagged XBRL from text.
Datamule Classifier AI 1 *B will be a fine-tuned model to cheaply identify AI generated text. May or may not work.
Datamule Classifier SEC Filings 1 *B will be a fine-tuned model to cheaply classify files as containing the information they are supposed to when submitted to the SEC.