datasets

You can build an ML dataset from JSONL files by constructing a SupervisedDatasetBuilder with a shared common_config, which automates data loading, normalization, and serialization for reproducible fine-tuning workflows.

The best way to prepare HuggingFace datasets for RLHF is to use an RLDatasetBuilder, which normalizes loaded data and serializes it into deployment-ready formats specifically structured for reinforcement learning pipelines.

You can convert conversation data into a training dataset using a ChatDatasetBuilder, which structures and normalizes conversational inputs into serialized formats suitable for supervised fine-tuning or evaluation workflows.

Yes, the dataset builder approach supports built-in configurations like Tulu3Builder and NoRobotsBuilder, allowing you to apply standardized preprocessing and normalization across diverse data sources through a shared common_config.

To serialize datasets across different formats, use the low-level helpers within the builders to convert, load, and normalize data from sources like JSONL or HuggingFace into reproducible outputs for your training pipelines.