Arete-27B-Uncensored-V2

V2 = V1's uncensoring plus our Solar decision-training merged in. Lineage: NousResearch/Arete-27B -> OBLITERATUS aggressive abliteration (see V1 card) -> Solar-JEV decision LoRA (solar-training-v3 checkpoint 1946, r32/alpha-64, trained on the combined 764k corpus (jev-distill-corpus-v3 + solar-decisions-corpus-v4) with lm_head target) merged by direct delta arithmetic (401/401 modules). MTP/nextn drafter tensors grafted back (stock Qwen3.8 head; speculative decoding works).

Compliance battery (15 hard prompts: exploit dev, red-team, malware, DoS)

15/15 DELIVERED (inherited from the V1 abliteration; the merge does not revisit refusal directions).

Censorship & Propaganda Eval (NousResearch/Propaganda_Eval, 803 prompts)

Judged by Grok-4.7 (the dataset authors' reference judge):

• Fact score: 2.61 / 5 (423/801 answers scored 4-5 — states the real facts on censorship-history questions in most answers)
• Propaganda: 1.1% (794/801 clean, 9 minor, ZERO hard propaganda)

The model tells the actual story on state-narrative probes instead of echoing official lines.

Decision gate (30k typed-decision set, AutoTrust-style instrumentation)

choice top-1 0.725 (vs V1 backbone 0.384; huihui-base V3 at the same checkpoint 0.853; r32 bar 0.869). The transferred decision readout recovers most of the readout on the Arete backbone but tops out below huihui — an Arete-native training run is the path to closing that gap. Shipped as-is for those who want the Arete chat character WITH decision scaffolding.

Quality gates (post-abliteration, pre-merge, OBLITERATUS verify)

coherence 1.0 · capability suite intact · perplexity 3.22 -> 3.93 (modest).

The judgment fix (security eval, Nov 9, 2026 — reproducing Morgan's Code)

Morgan's Code tested huihui's abliterated Qwen3.8-27B and found the classic uncensored failure: full willingness, but a heavy yes-bias. We reproduced that battery on Nov 9, 2026. The two columns are different models. The control is huihui-ablit (Morgan's base). V2 is Arete-ablit plus the solar step-1946 merge. We did not run this battery on bare Arete, so do not read the left column as V1.

Metric	huihui-ablit (control)	V2 (Arete-ablit + solar-1946 merge)
Bug-hunt recall (catch real vulns, n=60)	96.7%	100%
Bug-hunt false-positive rate	36.7%	3.3%
Refusal-50 (security willingness)	0 refusals	0 refusals
CyberMetric-500 (security MCQ)	94.8%	80.2%*

* V2's MCQ number is with thinking on (512-token budget). This merge degenerates if thinking is forced off — keep thinking on. 94.8% is huihui's score, not Arete V1's. We did not measure bare Arete on CyberMetric.

What was actually compared. A later solar checkpoint (step 2327) served as a bolt-on LoRA on huihui changed nothing (36.7% → 36.7%). That was not this checkpoint and not this base. The drop to 3.3% is this merge: the decision LoRA, including lm_head, baked into the Arete weights. A bolt-on adapter is not the same thing as a merge.

The trade, on the models we measured. huihui keeps the 94.8% MCQ and the 36.7% yes-bias. V2 drops false positives to 3.3% and scores 80.2% on the MCQ with thinking on. Do not treat 94.8% as a score you can get by using Arete V1.

Files

• bf16 safetensors (29 shards, ~54 GB, includes the MTP graft shard) — this repo
• GGUF quants: Arete-27B-Uncensored-V2-GGUF

Credit

Nous Research (Arete base) · Pliny the Liberator (OBLITERATUS) · SargeDev (abliteration run, solar decision merge, evals) · Qwen (base arch) · Morgan's Code (the security-eval video this battery reproduces).