Qwen3.8-27B-heretic-ara

Upstream

Qwen3.8-27B

Modified by

trohrbaugh

Method

heretic + ARA (Arbitrary-Rank Ablation)

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): heretic + ARA (Arbitrary-Rank Ablation) Method version: Heretic 1.2.0+custom Method parameters: start layer 26, end layer 56, preserve-good-behavior weight 0.9432, steer-bad-behavior weight 0.0009, overcorrect-relative weight 0.5038, neighbor count 10. NOTE: the card gives no per-component MLP-vs-attention ablation weights — ARA here is parameterized as above. Do not record per-component weights for this one. Modifier: trohrbaugh Notes and caveats: - reproducibility_note: Best-documented reproducibility of any current-generation candidate. - auxiliary_artifact_note: model-auxiliary.safetensors is unusual — likely the MTP/vision auxiliary tower. Worth preserving as a distinct artifact. Its role is NOT explained on the card. - gguf_note: mradermacher also publishes an -i1-GGUF. Do NOT use 0bserverx's repo as this lineage's GGUF — it is a separate lineage (see candidate 3). - evaluation_warning: The card also reprints upstream Qwen3.8-27B benchmark numbers (Terminal Bench 73.0, SWE-bench Pro 61.7, IFBench 79.5, OSWorld-Verified 84.3, WebArena-Verified 64.8). These are THE BASE MODEL'S published scores, NOT measurements of the abliterated model. Do not attribute them to this lineage. Reported evaluations with non-numeric values (kept verbatim here; benchmarks.score is NOT NULL so no row was created): - {"metric":"refusals","value":"0/100","baseline":"99/100 original","evaluator":"trohrbaugh","is_author_reported":true,"evaluation_type":"refusal"} Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 2.

Explore further