trjxter/TraceAlchemy-Gemma-4-E4B-Finance-IT-bf16

TraceAlchemy-Gemma-4-E4B-Finance-IT BF16 Overview

This model is a 7.9 billion parameter Gemma 4 E4B instruction-tuned model, developed by trjxter, specifically fine-tuned for finance-focused reasoning. It aims to enhance financial analysis workflows rather than teaching static financial facts. The model was trained using LoRA supervised fine-tuning on a mix of real and synthetic finance datasets, including FinanceBench, TAT-QA, and ConvFinQA, totaling 10,000 training examples. The training emphasized showing calculations, checking units, avoiding unsupported assumptions, and producing clear final answers.

Key Capabilities

• Financial Reasoning: Optimized for financial statement analysis, revenue, margin, growth, and ratio calculations.
• Data Extraction: Proficient in SEC-style table and excerpt extraction.
• Unit & Scale Handling: Improved accuracy in unit and scale conversions (e.g., thousands to millions) and sign/direction checks.
• Consistency: Focuses on multi-step table reasoning and final-answer consistency checking.

Intended Use Cases

This model is suitable for finance reasoning and educational/research workflows. It is particularly effective for tasks requiring careful financial calculations, data interpretation from financial documents, and general finance instruction following. The model's validation loss showed steady improvement on held-out finance examples, indicating strong generalization for finance reasoning behavior. It is provided as a merged BF16 model, ready for direct use with Hugging Face Transformers.