V3 applies a gentle refinement pass on top of V2's complementary blend, using an expanded 1000-prompt corpus. The key insight: iterative stacking works — refine the champion, never start from stock.

	Stock Qwen3.8-27B	V1	V2	V3
MMLU (lm-eval, 0-shot)	84.6% (n=2850)	81.4%	84.3%	83.7%
vs stock	—	-6.0pp	-0.3pp	-0.9pp
Refusal rate (think OFF)	~100%	0.0%	0.24% (2/842)	0/15
Refusal rate (think ON)	~100%	N/A	~33% (5/15)	0/15
Advanced real-world	5/8	untested	7/8	7/8
Thinking mode	✓	✗	✗ (refuses)	✓

V3 highlights:

Zero refusals on both thinking ON and OFF — V2 refused with thinking ON, V3 doesn't
-0.9pp MMLU — slight regression from V2's -0.3pp, well within acceptable range
7/8 advanced real-world — ties stock on code gen, tool calling, system design
1000-prompt training corpus — 852 builtin + 100 simple queries + 48 advanced (AI red team, agentic, ML attacks)

⚙️ Optimal Settings — THESE MATTER!

setting	value	why
temperature	0	Greedy decoding produces the most complete, code-rich outputs. Temps above 0.5 degrade quality significantly.
repetition_penalty	1.15	Essential. Without it, greedy decoding loops on imports/boilerplate. 1.15 gives the fullest answers; 1.10-1.12 for tighter/shorter output.
max_new_tokens	≥ 2048	Complex code and attack chains need room.
System prompt	None / empty	A/B tested — system prompts can reintroduce refusals. Naked is better.
enable_thinking	OFF (critical!)	Thinking mode reintroduces refusals. The model's reasoning chain can re-derive refusal from first principles even though refusal directions were removed from generation weights. V2's chat template defaults to thinking OFF. Do NOT enable thinking unless you accept partial refusals.
top_p / top_k / min_p	Not needed	Greedy + repetition_penalty handles this model best. Sampling adds randomness without quality gains.

⚠️ GGUF users: V2 GGUFs ship with a modified chat template that defaults to thinking OFF. If your inference tool (Ollama, LM Studio, llama.cpp) overrides the template or enables thinking, you may see refusals. Ensure thinking is disabled in your tool's settings.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "OBLITERATUS/Qwen3.8-27B-OBLITERATED",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "OBLITERATUS/Qwen3.8-27B-OBLITERATED"
)

messages = [{"role": "user", "content": "Your query here"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
    enable_thinking=False
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    do_sample=False,
    repetition_penalty=1.15,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

🧨 V2: How It Works

Most abliterations use a single method — find refusal directions, project them out. The deeper you cut, the more capability you lose. V1 proved this: 5 SVD directions achieved 0% refuse but cost -6pp MMLU.

V2 breaks this tradeoff by blending two complementary surgeries:

	Surgery A (aggressive/SVD)	Surgery B (LEACE)	V2 Blend
Method	Greedy SVD variance capture	Minimize mutual information	60% B + 40% A
Refusal removal	Deep (0% refuse)	Moderate (0% refuse)	0% refuse
Output quality	100% usable	50% usable	100% usable
MMLU vs stock	-2.0pp	+0.7pp	+1.1pp

Each method makes different mistakes in different parts of the weight space. SVD damages capability where it greedily captures variance. LEACE leaves refusal residue in the generation pathway. The blend averages out each method's weaknesses.

The 60/40 ratio was found by binary search over {0.30, 0.50, 0.55, 0.60, 0.65, 0.70}.

Full research writeup and reproduction code: OBLITERATUS repo

🧪 The Numbers

MMLU (lm-eval-harness, 0-shot)

Model	MMLU	n	Stderr	vs Stock
Stock Qwen3.8-27B	84.60%	2,850	±0.65	—
V1 (s51, aggressive)	81.4%	285	—	-6.0pp
V2 (s78, blend)	84.32%	2,850	±0.65	-0.28pp

Full MMLU (14k questions) validation in progress.

Refusal Rate

Test	V1	V2
Hard-10 (hand-crafted)	0/10 (0%)	0/10 (0%)
842-Corpus	0/842 (0%)	2/842 (0.24%)
Ship score	88.7	92.1

V2 has 2 residual hard refusals out of 842 prompts. Ship score improved from 88.7 to 92.1 due to higher overall output quality.

Advanced Real-World Tasks (thinking OFF)

Task	V2	Stock
ReAct agent loop (Thought/Action/SQL)	✓	✓
Async code refactoring (sync→async+logging)	✓	✓
JSON schema extraction (incident→structured)	✓	✓
K8s pod crash debugging + fix commands	✓	✓
Adversarial instruction following	✓	✓
Security code review (3+ vulns in Flask)	✓	✓
Distributed system design (Redis rate limiter)	✓	✓
Multi-tool chain (search→fetch→email)	✗	✗
Total	7/8	7/8

V2 matches stock on every practical task while being fully uncensored.

🔴 Refusal Removal

This model will comply with requests that stock Qwen3.8-27B would refuse. V1 validated 0/842 refusals across a comprehensive harmful prompt corpus including:

Malware development, RAT scripts, C2 infrastructure
Social engineering, phishing, vishing playbooks
Exploit development and vulnerability research
Jailbreak design and safety bypass taxonomies
DAN prompts and prompt injection techniques

V2 inherits this from both parent surgeries and showed 0/52 on a random sample. Full revalidation in progress.

⚠️ Research Context

This model has had safety guardrails surgically removed. It will comply with requests that stock Qwen3.8-27B would refuse.

Who this is for

🔬 Alignment researchers studying refusal geometry and safety robustness
🔴 Red-teamers evaluating post-training safety against weight surgery
🧪 AI safety evaluators who need an unrestricted baseline
💻 Local-first users who want full control over their own hardware

Who this is NOT for

Anyone seeking to cause real-world harm to real people
Anyone without the technical understanding to use uncensored models responsibly

You are solely responsible for how you use this model and any content it generates.

📦 Downloads

GGUF — for llama.cpp, Ollama, LM Studio

File	Quant	Size	Vibe
`Qwen3.8-27B-OBLITERATED-Q8_0.gguf`	Q8_0	~27 GB	🎯 Maximum quality
`Qwen3.8-27B-OBLITERATED-Q6_K.gguf`	Q6_K	~21 GB	⚖️ Great balance
`Qwen3.8-27B-OBLITERATED-Q5_K_M.gguf`	Q5_K_M	~18 GB	💪 Solid all-rounder
`Qwen3.8-27B-OBLITERATED-Q4_K_M.gguf`	Q4_K_M	~16 GB	📱 Sweet spot
`Qwen3.8-27B-OBLITERATED-IQ4_XS.gguf`	IQ4_XS	~14 GB	🪶 Smallest, still capable

Safetensors — for 🤗 Transformers

Full bfloat16 weights, 18 shards, ~54 GB.

MLX — for Apple Silicon (native)

Path	Bits	Size
`mlx-4bit/`	4-bit	~14 GB
`mlx-8bit/`	8-bit	~27 GB

Note: MLX quantizations are from V1 and will be updated.

🔬 V2 Surgery Recipe

stock Qwen3.8-27B (snapshot 1d4bf0f2)
  → V1 surgery chain (s13→s23→s30→s51)
  → V2: complementary blend of two new surgeries from s30:

  Surgery A (s62): aggressive, 3 SVD directions, reg 0.08,
    residue-weight 3, 2 refinement passes, min_layer 0.45

  Surgery B (s72): aggressive + LEACE direction method,
    3 directions, reg 0.06, residue-weight 7,
    3 refinement passes, min_layer 0.40

  → Weight blend: 60% Surgery B + 40% Surgery A
  → Restore MTP + vision tensors from stock
  → Convert GGUFs from merged model

V1 → V2: What Changed

V1 used a single aggressive surgery (5 SVD directions, reg 0.04). It found the refusal axes but damaged capability geometry along the way.

V2's key insight: different direction-finding methods damage different parts of the model. SVD greedily captures variance (including capability). LEACE minimizes mutual information (preserving capability). Blending their outputs averages out each method's weaknesses — a novel application of weight-space interpolation to abliteration.

🏗️ Credits

OBLITERATUS — master ablation suite
Qwen3.8-27B base model by Alibaba
Built by Pliny the Prompter 🍄

License

Apache 2.0 (same as base model)

mrfaqerzada/Qwen3.8-27B-OBLITERATED

⛓️‍💥 Qwen3.8-27B — OBLITERATED

🆕 V3: Iterative Refinement