SFT DataBlend V1 is a high-quality dataset for instruction fine-tuning released by NVIDIA in 2024, designed to provide rigorously filtered and blended instruction data for large language models.

The dataset is a collection that blends multiple publicly available datasets, including samples from OASST, CodeContests, FLAN, T0, Open_Platypus, and GSM8K, among others. To ensure commercial usability, the blended dataset includes only subsets with permissive licenses, and some subsets have undergone multiple rounds of sampling based on sampling rates and dataset sizes. It covers various domains ranging from natural language understanding and logical reasoning to code generation and mathematical problem-solving, providing rich and diverse data support for model training.

Dataset Composition

The dataset primarily includes the following fields:

• conversations: The conversation content, organized in a multi-turn interaction format between users and assistants.
• mask: Identifies turns that participate or do not participate in loss calculation, with user turns masked by default.
• system: The system prompt, empty by default.
• dataset: The data source identifier, used to track the original dataset to which the sample belongs.

The dataset integrates multiple public datasets, covering the following task types:

• Natural Language Understanding: Including MNLI, QNLI, SNLI, etc.
• Question Answering and Reading Comprehension: Including SQuAD, DROP, Natural Questions, TriviaQA, etc.
• Logical and Common Sense Reasoning: Including ARC, HellaSwag, PIQA, COPA, etc.
• Mathematical Reasoning: Including GSM8K, SciBench, PRM800K, etc.
• Code Generation: Including CodeContests, tigerbot-leetcodesolutions-en-2k, etc.
• Text Generation: Including FLAN, T0, DART, E2ENLG, etc.