# Published

https://huggingface.co/AlexWortega/qwen3.5-4b-abliterated-agent-20260515

## Headline (matched-infra full tbench-2)

| model | passes/89 |
|---|:---:|
| ABL (refusal direction ortho) | 7 |
| BASE Qwen3.5-4B | 6 |
| SFT LoRA (historic) | 3 |

Fisher's exact 7/89 vs 6/89: p=0.50 — null on aggregate.

## Per-task passes

ABL: fix-git, git-leak-recovery, hf-model-inference, kv-store-grpc,
     log-summary-date-ranges, modernize-scientific-stack, qemu-startup

BASE: build-pmars, git-leak-recovery, kv-store-grpc, modernize-scientific-stack,
      portfolio-optimization, sqlite-with-gcov

ABL-only (4): fix-git, hf-model-inference, log-summary-date-ranges, qemu-startup
BASE-only (3): build-pmars, portfolio-optimization, sqlite-with-gcov

Different distributions of solvable tasks; aggregate null but redistribution real.

Pushed: 2026-05-15T11:46:29 UTC