Rafaelcedav/atlas-r2-qwen3-14b

Rafaelcedav/atlas-r2-qwen3-14b: Forensic Reasoning Engine

This 14.7 billion parameter model, developed by Rafaelcedav, is a specialized fine-tune of Qwen3-14B designed for financial anomaly detection and forensic tax auditing in Mexico and the USA. It was trained in two distinct rounds on AMD Instinct MI300X hardware, focusing on progressively complex financial and legal datasets.

Key Capabilities & Differentiators

• Two-Round Specialized Training: Initial exposure to financial-legal vocabulary (Round 1) followed by deep refinement on high-complexity legal cases (Round 2), achieving a 15x improvement in loss.
• Explicit Reasoning ("Thinking Mode"): Leverages Qwen3's unique ability to articulate its thought process (<think>...</think>) before providing an answer, crucial for auditable legal contexts.
• Domain Expertise: Comprehensive knowledge of Mexican tax regulations (e.g., CFF, LISR articles) and US Internal Revenue Manual (e.g., IRM 4.10, BSA).
• Three-Level Forensic Reasoning: Operates from pattern detection (e.g., capacity vs. income, circular invoicing) to normative cross-referencing (mapping findings to specific articles, jurisdiction) and case construction (evidence synthesis, risk quantification).
• ATLAS Pipeline Integration: Designed as the core reasoning engine within a broader forensic audit pipeline, receiving extracted fields and generating traceable risk hypotheses.

Ideal Use Cases

• Automated Financial Auditing: Identifying potential tax evasion, money laundering, or financial fraud patterns.
• Legal Compliance Analysis: Cross-referencing financial data with specific tax laws and regulations in MX/USA.
• Risk Assessment: Quantifying estimated fiscal risk and suggesting prioritized corrective actions.
• Forensic Case Building: Generating auditable executive reports with synthesized evidence and legal arguments.