OpenAI MRCR — Multi-Round Co-Reference Resolution
Dataset containing long synthetic conversations where several identical queries (2, 4, or 8 occurrences) are hidden. The task is to extract the i-th specific occurrence requested, testing the fine management of the context by the models.
Around 800 examples structured in Parquet files, complex multi-turn dialogs
MIT
Description
OpenAI MRCR is a textual dataset dedicated to the resolution of multiple co-references over long contexts. It simulates multi-turn conversations where several identical requests are integrated, and the model must precisely identify the correct occurrence requested.
What is this dataset for?
- Evaluate the ability of models to manage long and complex contexts
- Testing accuracy in resolving multiple coreferences
- Advanced benchmarking for the robustness of dialogue models
Can it be enriched or improved?
It is possible to increase complexity by generating more examples with different numbers of occurrences, or to add dialogues from real cases to reinforce diversity. The annotation can be refined to include typical errors to be handled.
🔎 In summary
🧠 Recommended for
- NLP researchers
- Chatbot developers
- ML engineers specialized in dialogue
🔧 Compatible tools
- Pandas
- PyTorch
- Hugging Face Transformers
- Dialog tools
💡 Tip
Use textual similarity metrics to accurately assess responses in the multi-turn context.
Frequently Asked Questions
Is this dataset suitable for training conversational models?
Yes, it is designed to improve the management of long contexts in dialogue models.
How many examples does this dataset contain?
Approximately 800 examples divided into different context lengths and complexities.
Can custom variants of this dataset be generated?
Yes, by modifying the synthetic dialogues or by adding real cases to increase diversity.
