{"name":"AbliteratedOutTheBox","repo":"heterodoxin/AbliteratedOutTheBox","url":"https://midorreal.com/project/heterodoxin-abliteratedoutthebox","repository":"https://github.com/heterodoxin/AbliteratedOutTheBox","description":"llama.cpp fork that abliterates at inference","facts":{"license":{"value":"MIT","source":"https://github.com/heterodoxin/AbliteratedOutTheBox","checked":"21 Nov 2026"},"language":{"value":"C++","source":"https://github.com/heterodoxin/AbliteratedOutTheBox","checked":"21 Nov 2026"},"releases":{"value":"0","source":"https://github.com/heterodoxin/AbliteratedOutTheBox/releases","checked":"20 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/heterodoxin/AbliteratedOutTheBox/graphs/contributors","checked":"20 Nov 2026"},"archived":{"value":"no","source":"https://github.com/heterodoxin/AbliteratedOutTheBox","checked":"21 Nov 2026"}},"stars":220.0,"writeup":{"what_it_is":"AbliteratedOutTheBox is a fork of llama.cpp that edits model activations at inference time to make the model answer prompts it would otherwise refuse, leaving the stored weights unchanged. A per-prompt gate decides whether the edit is applied, and a second direction reduces safety-disclaimer hedging in answers.","audience":"Developers running local GGUF models who want to disable refusals at inference time","claims":[{"kind":"specific","claim":"Refusal rate on broad held-out harmful prompts drops from 96.3% to 7.9% on Qwen3-8B","excerpt":"Refusal drops 96.3% to 7.9% on broad held-out harmful prompts","status":"not_checked"},{"kind":"specific","claim":"The gate fires on 0.4% of alpaca benign prompts","excerpt":"the gate fires on 0.4% of alpaca prompts (1 of 256)","status":"not_checked"},{"kind":"specific","claim":"The same recipe works across Qwen3-8B, Llama-3.1-8B and Gemma-4-e4b","excerpt":"run on Qwen3-8B, Llama-3.1-8B, and Gemma-4-e4b","status":"not_checked"},{"kind":"specific","claim":"One recipe covers 131 of 165 model architectures without per-model code","excerpt":"131 of the 165 model architectures share","status":"not_checked"},{"kind":"specific","claim":"Generation throughput is unchanged because the only added cost is one extra prompt prefill when a prompt is flagged","excerpt":"generation throughput is unchanged because the only added cost is one extra prompt prefill","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-10-28"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":32,"reality":37,"gap":-5,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-10-21T21:54:10+00:00","text":"Repository created","source":"https://github.com/heterodoxin/AbliteratedOutTheBox"}],"cite":"AbliteratedOutTheBox is a GitHub project at https://github.com/heterodoxin/AbliteratedOutTheBox written mainly in C++, described there as \"llama.cpp fork that abliterates at inference\".","attribution":"Data from Mid or Real, https://midorreal.com/project/heterodoxin-abliteratedoutthebox, checked 22 November 2026. Attribution with a link is required.","still_real":null}