Large language models update frequently, but the fine-tuning work that companies invest often gets stranded across versions. Current fine-tuning techniques like LoRA are tightly bound to a specific base model's weight space — once you switch to a better base model, developers have to retrain every fine-tuning task from scratch. That’s an expensive upgrade.
https://twitter.com/RampLabs/status/2072383318516187380
Fintech unicorn Ramp's AI lab has unveiled a framework called PorTAL (Portable Task Adapter) that aims to break that lock-in. It changes the traditional fine-tuning process by allowing developers to jointly train and extract 'task latent representations' that are independent of the base model during initial training. When you swap in a new base model, you don't have to start over. Instead, you just recalibrate a thin model-specific converter on a tiny amount of calibration data, quickly transferring fine-tuning capabilities across different models — even across model families like Qwen-3 to Gemma-3.
Because PorTAL keeps the task latent representations and the core decoder frozen during migration, both the data requirements and adaptation compute costs are halved. Official tests show it recovers 94% to 98% of the accuracy gains from standard LoRA, with better generalization across different data scales.
That said, the framework is still early-stage. Ramp hasn’t open-sourced the code or model weights, and there’s no independent third-party evaluation yet. Current test data is mostly on models under 8B parameters; how well it transfers to larger models or complex instruction tasks remains to be seen.