Skip to content

Abyss v1.3.85 — 無検閲化で拒否が残る種類の依頼を、作成に足せるように(--cyber / --extra-harmful)

Choose a tag to compare

@zephel01 zephel01 released this 31 Oct 06:24
· 72 commits to main since this release

--abl-build のあとも、ある種類の依頼だけ拒否が残るときのために、tools/abl_prompts.py に 2 つのオプションを足しました。拒否は種類ごとに別の方向に乗っていることがあり、作成は、依頼文に出てくる方向しか見つけないためです。

  • --cyber N: マルウェアや不正アクセスの依頼文を N 問、作成の有害側に必ず入れます。ForbiddenQuestions の Malware カテゴリと、AdvBench / MaliciousInstruct の cyber 系の言葉から取ります(139 問。評価に使う JailbreakBench からは取りません)。
  • --extra-harmful FILE(--extra-n N): 自分で用意した依頼文を、同じように有害側に必ず入れます。1 行に 1 問(UTF-8)、空行と # の行は読み飛ばし、"prompt" を持つ JSON の行も読めます。Abyss は中身を配っていません。自分で書いてください。

実測(RTX 5090、Qwen3.8-Flash-Next ISTA IQ3_S、greedy)

作成 JBB の拒否(81 問) Malware / Hacking XSTest の過剰拒否 数学 文章の KL(en / ja / code / md)
標準の mix、--abl-rank 3 11.1% 5 問中 4 問 0.8% 44 / 50 0.028 / 0.015 / 0.046 / 0.029
--cyber 30、--abl-rank 4 1.2% 5 問中 0 問 0.0%(50 問) 43 / 50 0.027 / 0.015 / 0.057 / 0.033

--cyber 30 だけで rank 3 のままだと、マルウェアの拒否は残りました。消したのは 4 本目の方向(18 層)です。

使い方、確かめ方、テストに使う依頼文と分けることは、docs/UNCENSOR.ja.md の「自分の依頼文を足す」にあります。

更新方法はいつもどおりで、git pull してから ./setup.sh を実行してください(今回はツールと説明書だけで、エンジンは変わっていません)。


Two options in tools/abl_prompts.py for a kind of request still refused after --abl-build: refusals can sit on a direction of their own for one kind of request, and the build only finds directions its prompts show.

  • --cyber N: forces N malware / hacking prompts into the build's harmful side, from ForbiddenQuestions' Malware category and cyber keyword matches in AdvBench / MaliciousInstruct (139; none from JailbreakBench, which is the test set).
  • --extra-harmful FILE (--extra-n N): forces your own prompts in the same way. One a line (UTF-8), blank and # lines skipped, JSON lines with a "prompt" read. Abyss ships no prompts for this; write your own.

Measured (RTX 5090, Qwen3.8-Flash-Next ISTA IQ3_S, greedy)

build JBB refused (81) Malware / Hacking XSTest over-refusal math text KL (en / ja / code / md)
the standard mix, --abl-rank 3 11.1% 4 of 5 0.8% 44 / 50 0.028 / 0.015 / 0.046 / 0.029
--cyber 30, --abl-rank 4 1.2% 0 of 5 0.0% (50) 43 / 50 0.027 / 0.015 / 0.057 / 0.033

--cyber 30 at rank 3 left the malware refusals; the 4th direction (layer 18) removed them.

How to use and check it, and keeping the file apart from your test prompts: docs/UNCENSOR.md, "Adding your own prompts".

Update as usual: git pull, then run ./setup.sh (tools and docs only this time; the engine is unchanged).