{"name":"deleting-weights","repo":"agentic-learning-ai-lab/deleting-weights","url":"https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights","repository":"https://github.com/agentic-learning-ai-lab/deleting-weights","description":"Public code release for the paper \"The Surprising Effectiveness of Deleting Weights in LLM Post-Training\"","facts":{"license":{"value":"MIT","source":"https://github.com/agentic-learning-ai-lab/deleting-weights","checked":"22 Nov 2026"},"language":{"value":"Python","source":"https://github.com/agentic-learning-ai-lab/deleting-weights","checked":"22 Nov 2026"},"homepage":{"value":"https://agenticlearning.ai/deleting-weights/","source":"https://agenticlearning.ai/deleting-weights/","checked":"22 Nov 2026"},"releases":{"value":"0","source":"https://github.com/agentic-learning-ai-lab/deleting-weights/releases","checked":"21 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/agentic-learning-ai-lab/deleting-weights/graphs/contributors","checked":"21 Nov 2026"},"archived":{"value":"no","source":"https://github.com/agentic-learning-ai-lab/deleting-weights","checked":"22 Nov 2026"}},"stars":132.0,"writeup":{"what_it_is":"Official code for Deletion Tuning, a post-training method that learns a binary mask over the gate projection of transformer blocks in an LLM, zeroing a small fraction of pretrained weights. The repository covers training with GRPO, SFT, and on-policy self-distillation, saving masks, applying masks to base models, and evaluation.","audience":"Researchers working on LLM post-training and continual learning","claims":[{"kind":"specific","claim":"Zeroes fewer than 0.05% of weights and matches full fine-tuning at scale across three post-training regimes","excerpt":"zeroing fewer than 0.05% of its weights, with no other modification, matches full fine-tuning (FFT) at scale","status":"not_checked"},{"kind":"specific","claim":"Trained masks are hundreds to thousands of times smaller than the corresponding FFT checkpoint","excerpt":"the trained mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint","status":"not_checked"},{"kind":"specific","claim":"Forgets less than other methods, with prior-task accuracy staying near base level under reinforcement learning","excerpt":"its prior-task accuracy stays near the base level while FFT, LoRA, and random sparse fine-tuning lose up to 5.4 points","status":"not_checked"},{"kind":"broad","claim":"Method is described as learning which pretrained weights to delete","excerpt":"a post-training method that learns which pretrained weights of an LLM to delete","status":"too_broad"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-12"},"why_now":null,"owner_supplied":null,"score":{"verdict":"too_early","hype":56,"reality":18,"gap":38,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T07:00:00+00:00","verdict_source":"rule"},"early_read":{"estimate":true,"lean":"too_close","gap_low":-1,"gap_high":41,"rule":"early-2.0"},"changes":[{"date":"2026-11-10T16:19:07+00:00","text":"Repository created","source":"https://github.com/agentic-learning-ai-lab/deleting-weights"}],"cite":"deleting-weights is a GitHub project at https://github.com/agentic-learning-ai-lab/deleting-weights written mainly in Python, described there as \"Public code release for the paper \"The Surprising Effectiveness of Deleting Weights in LLM Post-Training\"\".","attribution":"Data from Mid or Real, https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights, checked 22 November 2026. Attribution with a link is required.","still_real":null}