remix3030303

cyber-qwen3-8b-agent-lora

README

License: apache-2.0

Training summary

Base model: Qwen/Qwen3-8B
Method: supervised fine-tuning (SFT) with LoRA on a 4-bit NF4 quantized base (QLoRA), assistant-only loss
Framework: TRL SFTTrainer, PEFT, bitsandbytes, Transformers
Hardware: 2x NVIDIA T4 (Kaggle), DDP via torchrun, fp16
Schedule: 1 epoch, 1073 optimizer steps, max sequence length capped at 2048 tokens
Final training loss: 0.874; token accuracy about 77 percent at the end of training
Checkpointing with resume was used because of CUDA out-of-memory restarts mid-run

Training data

A curated mix built by the training notebook: cybersecurity instruction data, grounded CTI (CVE, CWE, ATT&CK, CAPEC) records, a bounded sample of HolyCyber, tool-use trajectories (shell, http_request, search, read_file, grep, python), multi-turn investigations, ambiguity and false-positive examples, plus a small general-instruction mix to reduce catastrophic forgetting. Data was exact-deduplicated and split by entity so the same CVE/CWE/ATT&CK/CAPEC entity does not appear across train, validation and test. The dataset itself is not published here, and source licenses may differ, so check upstream terms before commercial use.

Intended use

Authorized penetration testing assistance, scoping and reporting
Triage of scanner output and false-positive analysis
Detection engineering, threat intelligence summarization, incident response notes
Secure code review and remediation advice
Agentic workflows that call tools within an authorized scope

Out of scope and limitations

Do not use against systems you do not own or have written authorization to test.
Not a replacement for a human security professional. Outputs can be wrong or outdated, so verify findings and CVE details independently.
No formal benchmark results are reported. The held-out sanity test and base-vs-finetuned benchmark were not run for this release.
Trained with sequences up to 2048 tokens; longer-context behavior is untested.
Inherits the biases and limitations of Qwen3-8B.

License

Apache-2.0, following the base model. Training data sources may carry their own terms.

This repo: LoRA adapter

PEFT LoRA adapter only. Load it on top of Qwen/Qwen3-8B.

python

import torchfrom transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype=torch.float16, device_map="auto")model = PeftModel.from_pretrained(base, "remix3030303/cyber-qwen3-8b-agent-lora")tok = AutoTokenizer.from_pretrained("remix3030303/cyber-qwen3-8b-agent-lora")

MLX 4-bit merged build for Apple Silicon: https://huggingface.co/remix3030303/cyber-qwen3-8b-agent-MLX-4bit

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

remix3030303

Model Tree

Base

Qwen/Qwen3-8B

Adapter

this model

Input Modalities

Text

Output Modalities