KBTG-Labs/THaLLE-0.1-7B-fa is a 7.6 billion parameter Qwen2-7B-Instruct model fine-tuned by KBTG Labs. This model specializes in financial knowledge, specifically excelling at CFA (Chartered Financial Analyst) mock exam questions. It demonstrates strong performance on internal and external CFA benchmarks, outperforming its base model and other 7B-8B class LLMs on these specific tasks. The model has a context length of 131072 tokens.
Loading preview...
THaLLE-0.1-7B-fa: Financial Expertise LLM
KBTG-Labs/THaLLE-0.1-7B-fa is a 7.6 billion parameter language model developed by KBTG Labs, built upon the Qwen2-7B-Instruct architecture. This model is specifically fine-tuned for financial domain knowledge, with a particular focus on CFA (Chartered Financial Analyst) exam questions.
Key Capabilities & Training
- Specialized Financial Knowledge: Fine-tuned on KBTG's internal CFA Mock Exam 2009-2019 dataset, comprising 9,426 questions, using LoRA.
- Enhanced CFA Performance: Demonstrates superior performance on various CFA-related benchmarks, including "Internal 2020", "Internal 2024", and "Flare CFA" datasets. It significantly outperforms its base model, Qwen2-7B-Instruct, and other models like Llama-2-7b-chat-hf, gemma-7b-it, and Meta-Llama-3-8B-Instruct on these financial tasks.
- Qwen2 Base: Leverages the robust Qwen2-7B-Instruct as its foundation, with a patched
bos_tokenin its tokenizer configuration for optimal performance.
Ideal Use Cases
- Financial Question Answering: Excels at answering multiple-choice questions related to CFA curriculum and financial analysis.
- Financial Education Tools: Suitable for developing tools for CFA exam preparation, financial knowledge assessment, or educational platforms.
- Domain-Specific Applications: Recommended for applications requiring high accuracy in understanding and responding to complex financial queries, particularly those structured similarly to exam questions.
For detailed results and methodology, refer to the Technical Report.