ramp-public/portallib

PorTAL, short for Portable Task Adapters for LLMs, is a Python research library built by Ramp Labs for training and sharing small, portable pieces of machine learning models called LoRA adapters. Normally, an adapter trained for one language model has to be retrained from scratch to work with a different base model. PorTAL tries to solve this by learning a shared representation of a task that is not tied to any single base model, plus a lightweight decoder for each specific base model that turns that shared representation into a standard LoRA adapter. In practice, this means a task can be trained once, then adapted to work with several different frozen language models, and finally exported as an ordinary Hugging Face PEFT adapter that works with the usual PyTorch and Hugging Face tools. The library, called portallib, handles loading, training, saving, publishing, and exporting these PorTAL adapters. The project ships with a set of ready made recipes that reproduce the training and evaluation approach described by Ramp Labs, along with several prebuilt adapters hosted on Hugging Face for models such as Qwen3 in 1.7 billion, 4 billion, and 8 billion parameter sizes, as well as Gemma 3. Installing the base library is a simple pip install, with an optional extra for the training dependencies and Hugging Face integrations. Python 3.11 and 3.12 are supported, and GPU training requires a CUDA compatible version of PyTorch. Beyond the plain Python workflows, the project also includes a command line interface driven by TOML configuration files, useful for running training, refitting, and evaluation jobs inside containers or scheduled automation. The CLI validates configuration files before running, reports progress as JSON lines, and reads authentication tokens from environment variables rather than the configuration files themselves, keeping credentials separate from recipe files. The README notes this is an alpha research release, so its APIs and file formats may still change before a stable version ships.