p-e-w/heretic
自動移除語言模型安全對齊的工具,保留原始能力降低拒答。 Fully automatic censorship removal for language models
abliterationllmtransformer
AI 推薦理由 Why AI picked it
以方向性消融與TPE參數優化自動搜尋低拒答且KL散度低的參數,支援多數dense、MoE與部分多模態模型;README提供基準比較與復現指令,適合能執行指令列、想自行產生去審查模型的使用者,不支援純狀態空間模型。 Uses directional ablation and TPE optimization to find parameters that minimize refusals and KL divergence, supporting dense, MoE, and some multimodal models. README includes benchmarks and reproduction commands. Suited for CLI users wanting an uncensored model; pure state-space models are not yet supported.
上榜紀錄 Trending history
- 2026-09-06 每週 Weekly #16
- 2026-09-05 每週 Weekly #11
- 2026-09-04 每週 Weekly #11
- 2026-09-03 每週 Weekly #13
- 2026-09-02 每週 Weekly #11
- 2026-09-01 每日 Daily #13
- 2026-09-01 每週 Weekly #19
- 2026-08-31 每日 Daily #5
- 2026-08-30 每日 Daily #6