ODA-Fin-RL-12K: 12,187 Verifiable Financial Reasoning Samples

OpenDataArena published ODA-Fin-RL-12K in March 2026, providing 12,187 hard-but-verifiable samples for reinforcement learning in the financial domain. The dataset focuses on complex reasoning tasks with concise answers optimized for automated reward modeling and distillation.

Use Cases

• Training Group Relative Policy Optimization (GRPO) models using the verifiable answer format
• Fine-tuning financial reasoning models on 'hard' samples to improve performance on complex domain-specific queries
• Evaluating the distillation of reasoning capabilities from larger models into smaller financial LLMs

Strengths

• 12,187 high-difficulty samples specifically curated for financial reasoning
• Reliably verifiable answers suitable for automated RL reward functions
• Bilingual support in English and Chinese (zh, en)

Limitations

• Small sample size of 12,187 records compared to typical SFT datasets
• Intentional difficulty bias excludes simple or common financial tasks

Provenance

Source

OpenDataArena (Arxiv: 260307223)

Collection Method

Curated distillation and difficulty-aware selection

Freshness

Last updated March 2026.