Shanghai AI Laboratory (OpenCompass), 2025
CNFinBench
Tests Chinese financial LLMs on expertise, agentic autonomy and integrity (safety and compliance).
About this benchmark
- Size
- 11,947 single-turn items over 26 tasks plus 321 adversarial dialogues; 22 models in the paper
- Public split
- Data and scripts on GitHub; the board uses a hard subset
- Grading
- Exact match for objective tasks, micro-F1 for extraction, three LLM judges with rubrics for open-ended tasks, and HICS for multi-turn safety.
- Licence
- Not stated
- Submission
- Through the OpenCompass leaderboard site. The paper does not say how to submit.
- Products allowed
- Models only
- Latest version
- v4 (arXiv, 2026-02-13)
- What it does not cover
- Non-Chinese markets and end-to-end products.
Results
No scores recorded here yet.
Papers and links
Last checked 6 Nov 2026. Something out of date or wrong? Send a correction.