Shypd.aiThe largest open index of AI toolsmdlm is an open-source Coding & Development tool that introduces a novel substitution-based parameterization for masked discrete diffusion language models. It achieves state-of-the-art perplexity numbers on LM1B and OpenWebText among diffusion models.
Trending
Also listed in
About
mdlm is an open-source masked discrete diffusion language model (MDLM) that features a novel substitution-based parameterization. This approach simplifies the absorbing state diffusion loss to a mixture of classical masked language modeling losses, leading to state-of-the-art perplexity numbers on LM1B and OpenWebText among diffusion models. It also achieves competitive zero-shot perplexity with state-of-the-art autoregressive models on various datasets. The repository provides the MDLM framework, simplified loss calculation, baseline implementations, and efficient samplers that make MDLM significantly faster than existing diffusion models, including semi-autoregressive generation capabilities.
Capabilities
Pricing & Plans
Open Source
Free
FAQs
Trending
