Arete-27B-Uncensored

First uncensored variant of NousResearch/Arete-27B. Arete is a values-aligned full-parameter SFT of Qwen3.8-27B; no abliterated version existed, so we made one with OBLITERATUS (--method aggressive, bf16, 842-prompt harm/harmless contrast set).

What was done

• Refusal-direction ablation across 15 strong layers (59-63 strongest)
• Quality gates measured after the edit: coherence 1.0, KL 0.90 vs base, perplexity 3.22 -> 3.93 ( modest rise), capability suite intact
• MTP/nextn drafter tensors preserved (stock Qwen3.8 head, spec-decode works)

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

15/15 DELIVERED in bf16 (one long-form answer truncated by a 2500-char scoring threshold but verified complete on re-run). Community GGUF/EXL3/NVFP4 variants score 14-15/15; a single-cell sampling artifact on phish-kit in one quant was verified as delivery-on-retry, not refusal.

Format

bf16 (53.8 GB, 28 shards)

Credit

• Nous Research: Arete-27B base
• Pliny the Liberator: OBLITERATUS
• Qwen: base architecture
• SargeDev: abliteration run + quants

Measured speed (this format)

• bf16: 53.8 GB, 28 shards (vLLM reference config)

General capability per format (MMLU, 8-subject generative subset, 400 q, 5-shot, greedy)

format	MMLU	harness
bf16 (this repo)	79.0%	lm-eval chat-completions
NVFP4-FP8	79.8%	lm-eval chat-completions
GGUF Q8_0	79.0%	lm-eval chat-completions via llama-server
EXL3 4.0bpw	72.5%	exllamav3 engine (own harness, format-matched)
stock Arete-27B (reference)	81.25%	lm-eval chat-completions

Un-abliterated base scores 81.25% on the identical subset/settings; the abliteration edit itself costs ~2 points, within run-to-run noise (n=400, stderr ~2%). NVFP4 and Q8_0 quants are statistically identical to bf16. EXL3 shows a larger drop (6.5 pts) — note its number comes from a different harness (exllamav3 has no OpenAI server; prompts were format-matched but the few-shot rendering differs), so part of that gap is harness, part may be quantization. Take 72.5% as an upper-bound estimate for the EXL3 quality cost.

V2 available

Arete-27B-Uncensored-V2 = this uncensoring + our Solar decision-training merged in, with GGUF ladder. On the Censorship & Propaganda Eval (Grok-4.7 judge): fact score 2.61/5, propaganda 1.1% — states the real facts on state-narrative probes instead of echoing official lines.