Overview
Venkat9990/finance-specialist-v7 is a 1.24 billion parameter Llama 3.2 Instruct model, developed by Naga Venkata Sai Chennu, specifically fine-tuned for finance-related conversations. A key design principle for this model was to prevent catastrophic forgetting, a common issue where fine-tuning for a specific domain degrades the model's general knowledge.
Key Differentiators & Technical Details
- Knowledge Preservation: Achieves minimal degradation of general knowledge (e.g., MMLU -0.19%, GSM8K -1.60%) compared to its base model, a significant improvement over previous versions (v1-v6).
- Targeted LoRA: Employs LoRA (r=8, alpha=16) with attention-only targets (q/k/v/o_proj), leaving MLP reasoning layers untouched to preserve core capabilities.
- Optimized Training: Utilizes a low learning rate (1e-5), rigorous data cleaning (72% of samples removed), and a single epoch to prevent overfitting and enhance stability.
- Base Model: Built upon the
unsloth/Llama-3.2-1B-Instruct architecture. - Training Data: Fine-tuned on the
Josephgflowers/Finance-Instruct-500k dataset, using 5,675 cleaned samples.
Performance Highlights
- General Knowledge: Benchmarks like MMLU, GSM8K, and IFEval show only minimal to moderate drops in performance compared to the base model, indicating successful knowledge preservation.
- Finance Domain: Demonstrates preserved or slightly improved performance on finance-specific MMLU benchmarks (e.g., Professional Accounting +0.35%).
- Recovery from Forgetting: Significantly recovers general knowledge and instruction following abilities compared to v6, with GSM8K improving by +25.92 points and MMLU by +7.19 points.
Use Cases
This model is ideal for applications requiring an AI assistant capable of engaging in financial discussions while retaining a strong foundation of general knowledge. It's particularly suited for scenarios where accuracy in financial information is critical, and the base model's broader reasoning capabilities must be maintained.