Data Tutorials

Choose the tutorial that matches how your examples should reach the training loop:

Workflow	Status and starting data	Tutorial	Runtime path
Pretraining from Megatron .bin/.idx	Recommended; prepared token IDs	DCLM preprocessing	GPTDatasetConfig → Megatron Core dataset builder
Text SFT or PEFT directly from Hugging Face	Recommended for hosted chat/prompt-completion rows or local JSON/JSONL loaded through Hugging Face datasets	Hugging Face text-only	HFDatasetSourceConfig → DirectHFSFTDatasetConfig → DirectHFSFTDatasetBuilder
Text SFT or PEFT from current prepared data	Transitional; materialized JSONL and optional packed Parquet	Prepared text-only	GPTSFTDatasetConfig → GPTSFTDatasetBuilder
Text SFT or PEFT from prepared .bin/.idx	Planned Issue #4664 workflow; not available yet	Follow Issue #4664	Future prepared-SFT builder
Multimodal SFT or PEFT from Hugging Face	Recommended for hosted vision, video, audio, and omni data or local JSON/JSONL loaded through Hugging Face datasets	Hugging Face multimodal	HFDatasetSourceConfig → DirectHFSFTDatasetConfig → DirectHFSFTDatasetBuilder
Large sharded multimodal training	Recommended for WebDataset/Energon data	Multimodal Energon	EnergonDatasetConfig → EnergonDatasetBuilder

Which SFT path should I use?

• Choose Hugging Face text-only for new on-the-fly text SFT, including hosted chat datasets, local conversation JSON/JSONL, and prompt-completion rows.
• Choose Hugging Face multimodal for processor-native image, video, audio, or omni conversations. Hosted datasets and local JSON/JSONL accepted by the Hugging Face json loader use the same HFDatasetSourceConfig; there is no separate preloaded/local VLM provider.
• Choose the current text-only SFT path when you specifically need reusable local JSONL, finite num_epochs, or offline packed-Parquet behavior. This is a transitional prepared-data path until the planned .bin/.idx SFT replacement is available.
• Choose the Energon tutorial for large sharded multimodal datasets; VALOR32K-AVQA is the production audio-video example.

All current SFT paths use serializable, declarative configuration and runtime builders. GPTSFTDatasetBuilder materializes Hugging Face text rows before constructing GPTSFTDataset; DirectHFSFTDatasetBuilder loads rows from Hugging Face datasets directly into DirectSFTDataset without intermediate JSONL materialization; EnergonDatasetBuilder constructs the configured task encoder and WebDataset loaders. The builders own runtime objects such as tokenizers, processors, task encoders, collators, and datasets.