Financial-RLVR-10K: 10,000 Financial Reasoning Dataset Released for GRPO Training

The Financial-RLVR-10K dataset has been released, designed for RLVR (Reinforcement Learning from Verifiable Rewards), GRPO, and PPO fine-tuning of open reasoning models such as Qwen, Llama, and DeepSeek.

Existing LLM-as-a-judge approaches are costly and can provide ambiguous reward signals. This dataset provides Python solution code for every problem, designed to enable deterministic rewards (1.0) through execution in a sandbox.

Key Features:

• 1,950 Adversarial Logic Traps: Accounting for 19.5% of the dataset, these train models to ignore invalid mathematical conditions (e.g., r <= g in the Gordon Growth model) and judge correct logic.
• Core Financial Domains: Includes DCF valuation, Black-Scholes option pricing, and corporate WACC.
• Open Source: Distributed under the MIT license, allowing free use by anyone.