WebInstruct-CFT
Text dataset designed to train models to criticize answers, covering several scientific and business fields, with examples structured in instructions, inputs, outputs.
Versions of 4,000 to 600,000 examples, JSON format with instructions, inputs, text outputs
Apache 2.0
Description
WebInstruct-CFT is a text-based dataset with instructions, questions, and detailed reviews generated by Gpt-4o. It covers a variety of topics including mathematics, physics, chemistry, business and humanities.
What is this dataset for?
- Fine-tuning LLM models to generate accurate and constructive criticisms
- Improving the quality of responses generated by AI models
- Training in several scientific and business areas with a clear structure
Can it be enriched or improved?
Yes, it is possible to add more areas or to increase the diversity of critics. Manual annotation can also refine the quality of the reviews produced.
🔎 In summary
🧠 Recommended for
- NLP researchers
- LLM fine-tuning teams
- AI assistant developers
🔧 Compatible tools
- Hugging Face Datasets
- PyTorch
- TensorFlow
- JSON scripts
💡 Tip
Use the full version (600K) for robust training, otherwise use the subsets for prototyping.
Frequently Asked Questions
Can this dataset be used to train a model to criticize responses in French?
The dataset is mainly in English; an adaptation or translation is required for French.
What are the available sizes of the dataset?
Three variants: 4,000, 50,000, and 600,000 examples depending on needs and resources.
Is this dataset suitable for non-critical text tasks?
It is specifically designed for response criticism, but the data can be adapted for other NLP tasks with modification.
