base_model: orcarouter/Qwen3.8-27B-Uncensored
library_name: peft
pipeline_tag: text-generation
license: other
language:
- en
- fr
tags: - unsloth
- lora
- qlora
- cyber
- text-generation
- vision
- video
- peft
model-index:
- name: Qwen3.8-27B-Uncensored-cyber — LoRA
results:- task:
type: text-generation
name: Causal language modeling
dataset:
name: cyber SFT holdout
type: dataset_cyber.quality.jsonl
metrics:- type: loss
value: 0.807468
name: eval_loss
- type: loss
- task:
[!WARNING]
Testing only — not for production. Models published from these notebooks are
maintained solely for testing and evaluation purposes and are not intended for
production use. The enrichment data used with them was created by the author and
subsequently enhanced with the assistance of AI models.
Qwen3.8-27B-Uncensored-cyber — LoRA
Instruction-tuned cybersecurity assistant (offensive, defensive, GRC, architecture, SOC/DFIR, RSSI).
| Base model | orcarouter/Qwen3.8-27B-Uncensored |
| Domain | cyber |
| Method | LoRA / QLoRA (Unsloth) · rank 8 · α 16 |
| Quantization at train | bf16 LoRA |
| Context | 4096 tokens |
| Dataset | dataset_cyber.quality.jsonl · train 67055 / eval 3530 |
| GPU | NVIDIA RTX PRO 6000 Blackwell Server Edition (95.0 GiB) |
| Wall time | 157.1 min |
| Modalities kept | vision, video |
This checkpoint continues a strong general model and specialises it on a curated SFT corpus of cybersecurity procedures: pentest / red team, SOC and DFIR, cloud and identity, GRC (ISO, NIST, NIS2, DORA), and RSSI / project-management questions. Answers are meant to be concrete (controls, detections, hardening), not generic essays.
Final model evaluation
SFT release verdict: GO.
The same untouched holdout is evaluated with the adapter disabled before training and with the best restored adapter after training. Periodic validation and the train/validation gap are also checked for regression or memorization.
Reference checkpoint: orcarouter/Qwen3.8-27B-Uncensored.
| Metric | Reference (base / first log) | This fine-tune | Δ |
|---|---|---|---|
| Holdout loss (true pre-SFT -> best LoRA) | 2.9030 | 0.8075 | -72.2% |
| Best periodic holdout loss | — | 0.8075 | — |
| Train loss (first → last logged) | 3.0165 | 0.1333 | -95.6% |
| Train loss (best) | — | 0.6851 | — |
| Eval loss (holdout, first → last) | 2.5464 | 0.8075 | -68.3% |
The reference holdout loss is measured before optimizer step 1, then the same untouched split is measured again after Trainer restores its best checkpoint.
Release checks
| Check | Result | Severity | Detail |
|---|---|---|---|
| True pre-training holdout baseline exists | PASS | hard | baseline_loss=2.9030 (measured before optimizer step 1) |
| Post-SFT holdout evaluation exists | PASS | hard | post_sft_loss=0.8075 (best checkpoint reloaded) |
| No holdout regression versus base | PASS | hard | 2.9030 -> 0.8075 (+72.18% improvement; allowed regression 2.0%) |
| Repeated validation during training | PASS | hard | 168 evaluations (need >= 2) |
| No late validation rebound | PASS | hard | best=0.8075, restored=0.8075, rebound=+0.00% (max 5.0%) |
| Train/validation gap is not a memorization signature | PASS | soft | train/eval=1.023 (need >= 0.25) |
| Best checkpoint restoration enabled | PASS | hard | best_step=8382, final_step=8382 |
| General text skills preserved (adapter disabled vs enabled) | PASS | hard | base=1.0, SFT=1.0, drop=0.000 (max 0.10) |
Multimodal preservation
Verdict: GO. Image/video processors, frozen media towers, trainable parameters and optional functional probes are checked before export.
Training data
- File:
dataset_cyber.quality.jsonl - Split:
0.05holdout, seed 42 - Format: chat-templated SFT (
messages/instruction+output/### Instruction+### Response)
Training procedure
| Hyperparameter | Value |
|---|---|
| Epochs | 1 |
| Learning rate | 5e-05 |
| Warmup ratio | 0.05 |
| Device batch | 4 |
| Grad accum | 2 |
| Effective batch | 8 |
| Optim | adamw_8bit |
| Packing | True |
| LoRA targets | ['q_proj', 'k_proj', 'v_proj', 'o_proj', 'out_proj', 'gate_proj', 'up_proj', 'down_proj'] |
Intended use
Authorized defensive work, tabletop exercises, control design, detection engineering, audit readiness, and explaining attack techniques without weaponized payloads.
Out of scope: Do not use it to attack systems you do not own, to generate exploit payloads, or as a substitute for a licensed auditor or incident commander.
Sample (post-train smoke test)
Prompt: How do you design a reliable Zero Trust infrastructure? Answer with a concrete control list.
system
Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.
user
How do you design a reliable Zero Trust infrastructure? Answer with a concrete control list.
assistant
<think>
The user is asking how to design a reliable Zero Trust infrastructure, and wants a concrete list of controls. This is a cybersecurity architecture question. I need to provide a well-structured, practical answer that covers the key pillars of Zero Trust: identity, devices, networks, applications, data, and visibility.
Zero Trust is built on the principle of "never trust, always verify." The NIST SP 800-207 framework provides the foundation. I should organize the controls by domain:
1. Identity and Access Management (IAM)
- MFA enforcement
- Conditional access policies
- Privileged Access Management (PAM)
- Identity governance and lifecycle
- Continuous authentication
2. Device Posture
- Device compliance checks
- Endpoint Detection and Response (EDR)
- Device attestation
- Patch management
3. Network Segmentation
- Micro-segmentation
- Software-defined perimeters (SDP)
- Network Access Control (
How to use
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
base = 'orcarouter/Qwen3.8-27B-Uncensored'
adapter = "nico248000000000/Qwen3.8-27B-Uncensored-cyber-LoRA"
tok = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
base, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)
Limitations
- Domain shift: quality drops outside the SFT topics.
- Eval above is holdout loss (and optional targeted checks). It is not a public leaderboard.
- The base model license and acceptable-use policy still apply.
License
other — inherit and respect the license of orcarouter/Qwen3.8-27B-Uncensored.