m3tl

Why we included this project

Training several related NLP tasks on one shared transformer backbone is the use case M3TL targets. It builds on HuggingFace transformers so a single model can handle classification, multi-label classification, sequence labeling, regression, masked language modeling, and contrastive learning together, which suits teams working on NER, part-of-speech tagging, and text classification over the same corpus. The library exposes the parts that usually make multi-task learning painful, like problem sampling, loss combination, and gradient surgery, as programmable modules, and includes several built-in multi-task algorithms so you can start without building the machinery yourself. Because it sits on top of transformers, you keep access to the usual pretrained model ecosystem. It is a research-oriented library rather than a maintained production service, so plan to adapt it to your own data pipelines, but for prototyping shared representations across NLP tasks it removes a lot of boilerplate.