The team behind DNA Factory says it has open-sourced its fine-tuning framework after using it internally for more than a year. It supports five training methods and includes a visualization of which tokens contribute to training loss.
Its creators say it is intended to make training behavior easier to inspect and reduce mistakes, including training on unintended tokens.
The team also reported results for an unreleased model called DNA 3.2, trained with domain knowledge injection. It says the model’s censorship score rose from 0.542 to 0.971 and its persona and domain-knowledge score from 0.076 to 0.855, with no drop on AIME, IFEval, MMLU-Pro, GPQA and other benchmarks; the post does not provide independent confirmation of these results.