NOTE: This repository is a work in progress. Changes and updates may occur as the project evolves.
Simplified and Stripped-Down ARC-Test-Time-Training (ARC-TTT). This implementation uses Hugging Face and unsloth, based on the original MARC repository, which used torchtune. Credits to MARC.
Create your conda environment for training with unsloth as explained here.
Does data augmentation on the training tasks and fine-tunes adapter per arc task and runs evaluation on the test task. Does not utilize a fine-tuned base model yet.
python ttt_unsloth.pyOriginal implementation uses a fine-tuned base model. In this repo we fine-tune task adapter from scratch of a regular pre-trained model that is not yet fine-tuned on the ARC tasks.
TTT performance over 20 tasks compared to untrained base model:
