本文へスキップ
Kitbay
スキル

obliteratus

MIT2026年11月11日 更新262,606GitHubスター

An optional Hermes Agent skill for OBLITERATUS, a separate AGPL-3.0 command-line tool that removes refusal behaviour, the built-in safety guardrails, from open-weight language models by editing their weights, a technique called abliteration, without retraining.

スキルの内容

An optional Hermes Agent skill for OBLITERATUS, a separate AGPL-3.0 command-line tool that removes refusal behaviour, the built-in safety guardrails, from open-weight language models by editing their weights, a technique called abliteration, without retraining. The agent installs the tool from its repository only after the user agrees, since it pulls in several gigabytes of dependencies, then uses its commands to inspect a model, get suggested methods, run the edit, compare methods and produce reports. Hardware needs grow with model size. The skill says to call the tool only through its CLI so its licence does not reach Hermes Agent, and to check perplexity and actual refusal rates afterwards, since edits can damage a model.

できること

  • Remove refusal behaviour from open-weight models
  • Suggest an abliteration method for a model
  • Compare methods and report results
  • Check edited models for quality loss