AI BriefingdNotitia open-sources DNA Factory, an opinionated LLM fine-tuning framework
·2026.11.10 20:35
Key point
The framework focuses on correct training by limiting options to SFT, DPO, GRPO, OPD, and SDFT, and includes a feature to visualize which tokens are included in the loss.
Details
- dNotitia open-sourced DNA Factory, a fine-tuning framework used internally for over a year, designed to make correct training the easy path by offering only one recommended approach for each task.
- The tool includes token-level loss-mask visualization using ANSI codes, where included tokens appear blue and excluded tokens appear gray, to prevent common errors where researchers train on unintended data.
- DNA Factory supports only five training methods: SFT, DPO, GRPO, OPD, and SDFT, with plans to remove DPO soon as it is rarely used; GRPO, OPD, and SDFT are on-policy methods aimed at avoiding catastrophic forgetting.
- A demonstration using the unreleased DNA 3.2 model, based on Qwen3.8 Flash Next, showed that domain knowledge injection increased censorship removal scores from 0.542 to 0.971 and new persona/knowledge scores from 0.076 to 0.855 without dropping performance on AIME, IFEval, MMLU-Pro, or GPQA.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
GitHub Next releases Ace, a local collaborative coding environment for humans and AI agents