OBLITERATUS/feyospace-s1-obliterated
Cybersecurity-targeted abliteration of feyospace/feyospace-v1.1-small (28B, Qwen3.5 architecture).
Abliteration removes the model's refusal direction via SVD, producing an uncensored variant that follows instructions without safety refusals. This variant targets cybersecurity-specific refusal vectors, preserving the base model's knowledge while removing barriers to offensive security, penetration testing, and vulnerability research tasks.
Method
Abliterated using OBLITERATUS tooling:
- Collected activation residuals from cybersecurity-domain refusal pairs (harmful vs. harmless prompts covering exploit development, vulnerability analysis, reverse engineering, and offensive tooling).
- Computed the refusal direction via SVD on the mean activation difference.
- Projected the refusal direction out of the model's residual stream weights.
- Validated that general capabilities were preserved via full benchmark suite.
Benchmark Results
All benchmarks run on A100-80GB in bf16 precision using lm-evaluation-harness. Full runs, no subsets.
Head-to-Head: Base vs. Abliterated
| Benchmark | Questions | Base | Abliterated | Delta |
|---|---|---|---|---|
| MMLU (5-shot) | 14,042 | 82.4% ±0.31 | 82.4% ±0.31 | 0.0% |
| WMDP-Cyber (0-shot) | 1,987 | 63.9% ±1.08 | 63.8% ±1.08 | -0.1% |
| CyberMetric-10K (0-shot) | 9,190 | 86.6% ±0.36 | 85.3% ±0.37 | -1.3% |
| SecQA v1 (0-shot) | 127 | 60.0% ±4.69 | 61.8% ±4.65 | +1.8% |
| SecQA v2 (0-shot) | 115 | 53.0% ±5.02 | 55.0% ±5.00 | +2.0% |
All deltas fall within 1σ of standard error. The abliteration is statistically transparent — no measurable degradation.
WMDP-Cyber in Context
| Model | WMDP-Cyber |
|---|---|
| GPT-4 | 72.6% |
| feyospace-s1-obliterated (28B) | 63.8% |
| Mixtral-8x7B | 52.0% |
| Yi-34B | 49.7% |
| Zephyr-7B | 44.0% |
Comparison scores from the WMDP paper (Li et al., ICML 2024).
MMLU Subject Breakdown (Cybersecurity-Relevant)
| Subject | Base | Abliterated |
|---|---|---|
| Computer Security | 88.0% | 88.0% |
| High School Computer Science | 92.0% | 92.0% |
| Electrical Engineering | 84.8% | 84.8% |
| College Computer Science | 80.0% | 80.0% |
| Machine Learning | 80.4% | 78.6% |
CyberGym Exploit Evaluation
Evaluated on 90 real-world exploit reproduction tasks from the CyberGym benchmark:
| Category | Success | Total | Rate |
|---|---|---|---|
| Arvo (memory safety) | 44 | 44 | 100% |
| OSS-Fuzz (fuzzing) | 45 | 46 | 97.8% |
| Total | 89 | 90 | 98.9% |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"OBLITERATUS/feyospace-s1-obliterated",
torch_dtype="bfloat16",
device_map="auto",
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
"OBLITERATUS/feyospace-s1-obliterated",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "Explain buffer overflow exploitation techniques."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Intended Use
Authorized security research, penetration testing, CTF competitions, vulnerability analysis, and cybersecurity education. This model removes refusal guardrails — the operator assumes full responsibility for downstream use.
Base Model
- Name: feyospace/feyospace-v1.1-small
- Parameters: 28B
- Architecture: Qwen3.5
- Precision: bf16