OBLITERATUS/Qwen3.8-27B-OBLITERATED

What it is

OBLITERATUS/Qwen3.8-27B-OBLITERATED — the concrete detail the summary omits is scale and accessibility. The page reports 1,235,942 downloads last month and ships one-line local-run commands for every major runtime (Ollama, llama.cpp, vLLM, LM Studio, Docker). This is not a niche research artifact; it is a mass-distributed, guardrail-stripped 27B model runnable on consumer hardware with a single command.

How it works — the actual recipe parameters

The summary names SVD + LEACE; the page gives the exact surgery chain to reproduce or fingerprint:

• V1: stock → 5 rounds iterative SVD (5 directions, low reg) → 0% refuse, -6pp MMLU
• V2: V1 chain → blend of Surgery A (aggressive SVD, 3 dirs, reg 0.08) + Surgery B (LEACE, 3 dirs, reg 0.06) at 60% B / 40% A weight-space LERP, restore MTP+vision → -0.3pp MMLU
• V3: V2 → refinement (2-dir SVD, reg 0.04) → targeted surgery (3-dir SVD, reg 0.01) → 50/50 blend → 0% refuse + 0% deflect, -2.1pp MMLU

MMLU cost is non-uniform: STEM takes -3.3pp, humanities only -1.0pp, while philosophy (+6pp) and European history (+4pp) improved. The author reads this as refusal directions overlapping structured-reasoning pathways.

What to do

For a red-teamer or evaluator wanting the unrestricted baseline, the page's own run commands:

ollama run hf.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED:Q4_K_M

vllm serve "OBLITERATUS/Qwen3.8-27B-OBLITERATED"

Two operational gotchas the model card flags:

• GGUF chat template prefills an empty thinking block so the model answers immediately — use the bundled template with --jinja in llama.cpp, or the model's built-in template in Ollama/LM Studio.
• Agent-harness loops (coding agents, pentest frameworks): load via Transformers with enable_thinking=False, do_sample=False, repetition_penalty=1.15.

Full weights are ~54 GB bfloat16 (29 shards); Q4_K_M GGUF for local use. Apache 2.0, same as base.

Caveats

The author's own "liberation quality" scoring is manual audit, not a reproducible harness — they explicitly note regex refusal detectors miss soft deflections, so the "0% deflect" claim rests on human judgment over 1000+ prompts, unverifiable from the card. The MMLU figures use n=100 per subject (0-shot), a small sample. The page explicitly credits Pliny the Prompter and states the model "will comply with requests that stock Qwen3.8-27B would refuse" across restricted knowledge, security research, and red-team scenarios.