naazimsnh02/fraudsentinel-qwen3-14b-merged

TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 9, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The naazimsnh02/fraudsentinel-qwen3-14b-merged model is a 14 billion parameter Qwen3-based language model, fine-tuned for enterprise fraud detection and AML investigation. It excels at structured JSON risk scoring, typology classification, and Suspicious Activity Report (SAR) drafting, offering a specialized intelligence layer for financial crime analysis. This model integrates merged LoRA adapter weights, enabling direct deployment without PEFT overhead for tasks like explainable alerts and multi-turn human-in-the-loop dialogue.

Loading preview...

FraudSentinel — Qwen3-14B Merged

This model, developed by naazimsnh02, is a 14 billion parameter Qwen3-based language model with merged LoRA adapter weights, designed as a specialized Tier-2 intelligence layer for enterprise fraud detection and AML investigation. It is ready for direct deployment without PEFT or adapter management.

Key Capabilities

  • Structured JSON Risk Scoring: Provides calibrated scores (0.0–1.0), risk levels, typologies, key signals, feature importance, and recommended actions, including SAR rationale.
  • Explainable Alerts: Generates evidence-grounded explanations for investigators, linked to transaction features.
  • Typology Classification: Identifies various fraud types such as card-not-present, account takeover, structuring, and smurfing.
  • 6-Level Recommended Actions: Suggests actions ranging from AUTO_APPROVE to SAR_REVIEW.
  • SAR Drafting: Generates FinCEN-aligned Suspicious Activity Report narratives for human review.
  • Multi-turn HITL Dialogue: Supports investigator follow-up conversations.
  • Deep Analysis Mode: Utilizes Qwen3's thinking tokens for Chain-of-Thought reasoning, adding ~3–5 seconds latency.

Training Details

Fine-tuned on unsloth/Qwen3-14B using SFT + LoRA, with 64 million trainable parameters (0.433% of the base model). It was trained on the naazimsnh02/fraud-financial-crime-qwen3-sft-v2 dataset (11,016 examples) for 2 epochs, achieving a final train loss of 0.2467. The model supports a maximum sequence length of 4,096 tokens.

Limitations

This model is intended for prototype/research use with synthetic/semi-synthetic data. It requires independent validation and human-in-the-loop controls for real-world adjudication. AI-generated SAR drafts need human review before submission. Full-precision inference requires a GPU with at least 40 GB VRAM, though 4-bit quantization can reduce this to ~10 GB.