rLLM/rLLM-FinQA-4B
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Jan 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm
rLLM/rLLM-FinQA-4B is a 4 billion parameter financial question-answering agent developed by rLLM, fine-tuned from Qwen3-4B-Instruct-2507 using reinforcement learning. This model is specialized in answering questions about SEC 10-K financial statements by utilizing tools like SQL queries, table lookups, and calculators. It achieves 59.70% accuracy on the Snorkel Finance Benchmark, making it highly effective for financial analysis tasks.
Loading preview...
FinQA: Financial Question-Answering Agent
rLLM/rLLM-FinQA-4B is a 4 billion parameter model developed by rLLM, specifically designed for financial question-answering. It is fine-tuned from Qwen3-4B-Instruct-2507 using Reinforcement Learning (RL) with LLM-as-judge rewards for correctness evaluation.
Key Capabilities
- Financial Statement Analysis: Answers questions directly from SEC 10-K financial statements.
- Tool Use: Integrates specialized tools for financial analysis, including:
get_table_names: Lists available tables for a company.get_table_info: Retrieves table metadata, columns, and sample values.sql_query: Executes SQLite queries on financial tables.calculator: Evaluates mathematical expressions.
- Data Handling: Trained on a dataset of 5,110 question-answer pairs derived from 207 companies' SEC 10-K filings, covering both single-table and multi-table reasoning.
Performance
- Achieves 59.70% accuracy on the Snorkel Finance Benchmark.
- Achieves 26.6% accuracy on Snorkel Finance Reasoning, matching gpt-5-nano-2025-08-07.
- Outperforms its base model, Qwen3-4B-Instruct-2507, which scored 27.90% on FinQA and 13.90% on FinQA Reasoning.
Good For
- Automated financial analysis and reporting.
- Extracting and reasoning over data from SEC 10-K filings.
- Applications requiring precise, tool-augmented responses to financial queries.