OpenReward is a repository of shared reward functions for language modelling tasks.

• Basic structure of the project
• Reward function for formatting
• Reward function for math problem correctness
• Ability to integrate with open-r1 project or huggingface TRL grpo demo

• Data loading - allow reward functions to be paired with datasets that exercise the reward function
• Code syntax reward function
• Code correctness reward function
• Code speed reward function
• Human feedback reward function (offline)
• Neural network reward function
• Language consistency reward function
• Function calling format reward function

Q: Where do we decide what reward functions are run for a given generation? Do we decide this at all? A: For now, just write the reward functions and let the user decide which ones to run.

Q: Is there a chat format we should use for the inputs to the reward function? A: We should use the messages/chat format from the OpenAI API.

OpenReward has built in support to integrate nicely with the following projects:

• TinyZero
• Huggingface TRL GRPO/open-r1