by OpenBMB
UltraData-SFT-2605 is OpenBMB's set of over 15 million supervised fine-tuning samples used to post-train a small on-device language model. It covers math, code, knowledge and instruction following, with both deep-thinking and non-thinking data.
DatasetFree
Step 1: Sign in to Hugging Face and accept the access conditions on the dataset page.
Step 2: Install the datasets library.
Step 3: Load the config for the domain you want, such as math or code.
Step 4: Choose the think or no_think data depending on whether you want reasoning traces.
Step 5: Mix the subsets into your SFT run and hold out your own evaluation set.
Apache 2.0. You still need to accept the access conditions on Hugging Face.
Deep Thinking data targets reasoning, planning and verification on complex tasks. Non-thinking data targets fast, direct answers.
It goes through query filtering, answer quality filtering, validation training and decontamination against existing benchmarks.