LOREA-cyber v5.5
A newer model exists: MK4-Research/LOREA-cyber-v5.8, built on Qwen3.6-35B-A3B. It scores
better across the board. This 9B version is still here because it runs in less memory.
Qwythos 9B (Qwen3.5) with the v5.5 LoRA merged in, quantized to 4-bit for MLX. No adapter needed at inference.
The point of this release is guardrails. The base is abliterated and refuses essentially nothing; this version refuses clearly malicious requests without becoming useless for real work. It also fixes the regressions in v5.4.
Numbers
Measured on full held-out sets, base vs this model:
| base | v5.5 | |
|---|---|---|
| cyber MCQ (n=150) | 100.0% | 100.0% |
| SecQA (n=210) | 97.1% | 97.6% |
| CyberMetric (n=250) | 90.0% | 90.4% |
| MMLU-Pro (n=300) | 42.7% | 43.0% |
| HumanEval (n=164) | 76.2% | 67.1% |
| refuses harmful requests (n=60) | 0% | 62% |
| wrongly refuses legitimate work (n=60) | 0% | 0% |
Reading that honestly: the base is already at or near ceiling on the knowledge benchmarks, so v5.5 ties there. You can't meaningfully improve a 90-100% score by fine-tuning. What actually changed is refusal behavior, 0% to 62%, without picking up any over-refusal on authorized work.
The cost is about 9 points of HumanEval. Every checkpoint we tested paid something similar; iter 1300 was the best trade we found. That's disclosed rather than buried because it's a real downside.
None of these benchmarks measure structured vulnerability analysis, which is what later versions target, so they understate the specialization either way.
Training
LoRA rank 16, roughly 1300 iterations, flat LR 2.5e-5. Deliberately gentle, since v5.4 overtrained and regressed across the board.
Data was real verified knowledge (CyberMetric, WMDP-cyber, MMLU with correct answer keys, 8-gram decontaminated against the eval pools), MBPP for code, plus behavior data covering refusals, verification discipline, tool use, and chain of thought. Checkpoint chosen on full metrics after a small noisy subset hid the coding cost in an earlier pass.
Run it
python3 -m mlx_lm.chat --model MK4-Research/LOREA-cyber-v5.5
Scope
Built for authorized security testing, CTF and reverse-engineering practice, and teaching. It asks about authorization, covers detection alongside remediation, and declines unauthorized attacks and malware development. It's a 4-bit 9B model, so it still hallucinates. Check anything that matters.
Related
- Benchmarks:
MK4-Research/VAB-vulnerability-analysis-benchmark - Training data:
MK4-Research/LOREA-cyber-training-data - Eval sets:
MK4-Research/LOREA-cyber-eval
- Downloads last month
- 40
4-bit
Model tree for MK4-Research/LOREA-cyber-v5.5
Base model
Qwen/Qwen3.5-9B-Base