Karthik-sr/qwen-atlas-raft-v1

Qwen-ATLAS LoRA Adapter

QLoRA adapter fine-tuned on Qwen 2.5 7B Instruct using Retrieval-Augmented Fine-Tuning (RAFT) on the MITRE ATT&CK STIX v2.1 knowledge base.

Model Description

This adapter conditions the base model to reason over retrieved ATT&CK context documents rather than relying on parametric memory. It is designed to be used with a ChromaDB RAG pipeline over the MITRE ATT&CK enterprise dataset.

This adapter is not useful without the retrieval pipeline.

Training

Parameter	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	QLoRA (4-bit NF4)
LoRA rank	16
Target modules	q_proj, k_proj, v_proj, o_proj
Training examples	1,743 RAFT samples
Epochs	2
Dataset	MITRE ATT&CK STIX v2.1

Evaluation (with RAG)

Configuration	Score
Base Qwen 2.5 7B, no RAG	35/80 (43.75%)
RAFT adapter, no RAG	12/80 (15.00%)
Base Qwen 2.5 7B + RAG	67/80 (83.75%)
RAFT adapter + RAG	59/80 (73.75%)

The 12/80 without RAG is expected and intentional — the model was trained to depend on retrieval context, not memorize ATT&CK facts.

Intended Use

Threat intelligence queries grounded in MITRE ATT&CK:

• Technique attribution and explanation
• Threat actor TTP profiling
• Tactic-filtered group queries
• Multi-hop ATT&CK relationship analysis

Project

Part of Qwen-ATLAS — an adversarial security research project studying retrieval poisoning and RAG system vulnerabilities in threat intelligence contexts.