Rafaelcedav/atlas-finanzas-deepseek-r1-8b
Rafaelcedav/atlas-finanzas-deepseek-r1-8b is an 8.3 billion parameter model, distilled from a 671B DeepSeek-R1 model, and fine-tuned for forensic financial and tax audit anomaly detection in Mexico and the USA. Developed by Rafaelcedav as part of the ATLAS system, it leverages DeepSeek-R1's pure reinforcement learning-based reasoning capabilities to generate structured chains of thought. This model excels at auditable reasoning for complex financial and legal cases, specifically focusing on tax regulations like Mexico's CFF/LISR and USA's IRS/IRM.
Loading preview...
ATLAS DeepSeek-R1 Finanzas: Forensic Tax Reasoning
This model, Rafaelcedav/atlas-finanzas-deepseek-r1-8b, is an 8.3 billion parameter fine-tuned version of DeepSeek-R1-Distill-Llama-8B, specialized in forensic financial and tax audit anomaly detection for Mexico and the USA. It was developed by Rafaelcedav as part of the ATLAS system during the AMD Hackathon 2025, leveraging AMD Instinct MI300X hardware.
Key Differentiators & Capabilities
- Distilled Reasoning: Unlike most LLMs that learn to respond, this model is distilled from a 671B DeepSeek-R1 model trained with pure reinforcement learning, enabling it to reason by generating structured chains of thought before providing answers. This preserves advanced reasoning capabilities in a significantly smaller and more efficient package.
- Auditable Conclusions: For forensic auditing, this translates to explicitly evaluated alternative hypotheses, evidence weighting, auditable and reversible reasoning, and conclusions with argumentative traceability.
- Specialized Domain Knowledge: Extensive training on financial and legal datasets for both Mexico (CFF, LISR, RMF RESICO) and the USA (IRS, IRM, BSA/FBAR, Common Law Test, Economic Reality Test).
- Efficient Training: Underwent multi-round fine-tuning on AMD MI300X, achieving rapid convergence (e.g., 25 minutes for legal specialization) due to the distilled Llama architecture's efficiency with bfloat16 native processing.
Ideal Use Cases
- Forensic Financial Auditing: Identifying anomalies and constructing arguments in complex financial investigations.
- Tax Compliance & Analysis: Analyzing specific tax articles, forensic procedures, and cross-border MX/USA tax implications.
- Automated Reasoning Systems: Serving as a core component in pipelines requiring explicit, traceable reasoning for high-stakes financial and legal decisions, as demonstrated in the ATLAS pipeline for processing vision-based findings into auditable verdicts.