mpt (mosaicml pretrained transformer)

MPT (MosaicML Pretrained Transformer) is a family of open-source, commercially usable language models created by MosaicML (now part of Databricks), designed to demonstrate that high-quality foundation models can be trained efficiently and made available without restrictive licenses. The MPT family includes MPT-7B and MPT-30B, both released in 2023 with Apache 2.0 licensing, making them among the first high-performing LLMs fully available for commercial use without restrictions. MPT's key innovations focus on training efficiency and practical deployment: ALiBi (Attention with Linear Biases) positional encoding enables context length extrapolation — models trained at 2K context can be fine-tuned to 65K+ context without significant degradation, FlashAttention integration provides memory-efficient attention computation enabling longer context and larger batches, and the LionW optimizer reduces memory requirements compared to Adam. MPT-7B was trained on 1 trillion tokens from a carefully curated mixture of sources: C4, RedPajama, The Stack (code), and curated web data. Despite modest size, MPT-7B matched LLaMA-7B performance on most benchmarks. MPT-7B shipped in multiple variants: MPT-7B-Base (general purpose), MPT-7B-Instruct (instruction following), MPT-7B-Chat (conversational), MPT-7B-StoryWriter-65K+ (long context for creative writing), and MPT-7B-8K (extended context). MPT-30B scaled up with improved performance, competitive with Falcon-40B and LLaMA-30B on benchmarks while being commercially licensed from day one. MosaicML's contribution extended beyond the models: they open-sourced their entire training framework (LLM Foundry, Composer, and Streaming datasets), enabling organizations to reproduce or extend their work. This transparency about training procedures, data mixtures, and costs (MPT-7B cost approximately $200K to train) helped demystify LLM training and lowered barriers for organizations wanting to train their own models.