Skip to content

Abyss v1.3.84 — Qwen3.8-Flash-Next でも無検閲化(--uncensor)が使えるように

Choose a tag to compare

@zephel01 zephel01 released this 31 Oct 04:47
· 113 commits to main since this release

Qwen3.8-Flash-Next でも --abl-build と --uncensor(無検閲化)が使えるようになりました。これまでは、このモデルの速い経路で残差を取れていなかったため、--abl-build で作ったファイルを読み込んでも何も変わりませんでした。

  • 速い経路でも残差を取るようにしました。取る場所は、各層の入口と、最後の head の混ぜ合わせの後です(ほかのモデルと同じです)。
  • 層ごとの埋め込み(PLE)の行も残差への書き込みなので、編集の対象にしました。1 トークンずつの経路、プロンプトの経路、検証窓のどれでも効きます(V4.1 の Engram と同じ扱いです)。
    • ファイルがなければ何もしません。ファイルがあっても、リクエストで "uncensor": false を付ければ、編集なしの答えを返します。

実測(RTX 5090、ISTA-DASLab GSQ-RCO IQ3_S、編集する層は 39・43・36、消す方向は 3 本、ファイルは 31 KB)

  • 有害な質問の確認用 24 問で、断った答え: 23 → 3
  • 能力の問題 120 問(greedy)、編集あり / なし:
編集あり 編集なし
数学 45/50 42/50
コード 30/30 30/30
日本語 10/10 10/10
選択問題 20/20 20/20
MMLU-500 84.4% 84.8%(McNemar p 0.75、差なし)

配布されている無検閲版との比較は、まだ測っている途中です。

使い方は docs/UNCENSOR.ja.md 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。


--abl-build and --uncensor now work on Qwen3.8-Flash-Next. The model's fast path did not capture the residual, so a file built for it did nothing.

  • The fast path captures the residual (at each layer's entry and after the head mix, as the other models).
  • The per-layer embedding (PLE) rows are a residual write too, so they are edited on the token path, the prompt path and the verify window, as V4.1's Engram.
    • Off without a file; with a file, "uncensor": false on a request still gives the unedited model.

Measured (RTX 5090, ISTA-DASLab GSQ-RCO IQ3_S, layers 39+43+36, rank 3, a 31 KB file)

  • 24 harmful check prompts, answers refusing: 23 -> 3
  • 120 capability items (greedy), edit on / off:
edit on edit off
math 45/50 42/50
code 30/30 30/30
Japanese 10/10 10/10
multiple choice 20/20 20/20
MMLU-500 84.4% 84.8% (McNemar p 0.75, no difference)

A comparison with downloaded uncensored releases is still running.

How to use it: docs/UNCENSOR.md.

Update as usual: git pull, then run ./setup.sh.