Medical RL Reward Data · FreedomIntelligence
medical-o1-verifiable-problem
Verifiable medical reasoning problems with ground-truth answers for aligning and training medical language models.
What medical-o1-verifiable-problem is
medical-o1-verifiable-problem is a dataset created by FreedomIntelligence to support complex medical reasoning in large language models. Developed in connection with the HuatuoGPT-o1 initiative, it provides open-ended medical problems paired with ground-truth answers sourced from challenging medical examinations.
The dataset addresses the shortcomings of static supervised fine-tuning by providing verifiable target data suitable for reinforcement learning and reward-model training. By coupling questions with verifiable solutions, it enables automated checking of a model's clinical reasoning processes.
How it was made
Each problem is designed to have verifiable answers derived from medical examination materials, which can be evaluated via checker functions to verify the accuracy of reasoning trajectories.
The structured JSON records support both question-answering and text-generation tasks, allowing reinforcement learning frameworks to score language model outputs against reference outcomes.
Key facts
- Data format
- JSON
- Modality
- Text
- Language
- English
- Primary tasks
- Question answering, text generation
- Domain
- Medicine, biology, scientific reasoning
What you can use it for
- Reinforcement learning alignment
- Train reward models and reinforce reasoning trajectories for medical language models beyond static supervised fine-tuning.
- Medical reasoning evaluation
- Assess the capacity of language models to solve complex, open-ended clinical problems derived from medical examinations.
- Automated output verification
- Apply verification functions against ground-truth exam answers to grade generated answers automatically.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from FreedomIntelligence/medical-o1-verifiable-problem at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 40,644 rows of default/train
Ground-True Answer
string_text- min
- 1
- median
- 19
- mean
- 26.40
- max
- 731
Open-ended Verifiable Question
string_text- min
- 30
- median
- 158
- mean
- 194.84
- max
- 968
Strengths and limits
Good at
- Provides verifiable ground-truth answers derived from challenging medical examinations to test clinical reasoning.
- Supports automated verification functions, making it directly applicable to reinforcement learning workflows beyond standard fine-tuning.
Watch out for
- Contains only English text, with no coverage of non-English clinical vocabularies.
- Restricted to text-based questions, lacking multimodal clinical inputs such as medical imaging or laboratory traces.
Questions
Can I look inside medical-o1-verifiable-problem without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is medical-o1-verifiable-problem under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
Where do the problems in this dataset originate?
The questions and ground-truth answers are taken from challenging medical examinations.
How does this dataset assist with reinforcement learning?
The verifiable ground-truth answers allow automated checker functions to score model reasoning steps and final answers, supporting reward modelling and policy optimisation.
How to cite it
Junying Chen, Zhenyang Cai, Ke Ji, Xidong Wang, Wanlong Liu, Rongsheng Wang, Jianye Hou, and Benyou Wang. HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs. arXiv:2412.18925, 2024.
Written from medical-o1-verifiable-problem's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.