introml-clara-implementation

Course project repo for an Intro to ML class: research + hands-on implementation of CLaRa (Continuous Latent Reasoning) and small-scale experiments on Kaggle T4.

• Reimplementation (Stage I/II training) and dataset loaders.
• Apple-native evaluation and fine-tuning using their original modeling_clara.py (with modifications for quantization and VRAM optimization) with a Kaggle-ready workflow.
• Kaggle notebooks for reproduction (train from scratch, fine-tune, evaluate).

Kaggle notebooks

Open any notebook in the notebook/ folder on Kaggle. The notebooks handle cloning, deps, and environment patching.

Main notebooks:

• notebook/clara-ml-final-evaluation-apple.ipynb — Apple-native eval on SQuAD + TriviaQA.
• notebook/clara-ft-squad.ipynb / notebook/clara-ft-triviaqa.ipynb — fine-tune on SQuAD or TriviaQA.
• notebook/clara-eval-ft-squad.ipynb / notebook/clara-eval-ft-triviaqa.ipynb — evaluate fine-tuned checkpoints.

Checkpoints as Kaggle datasets

The notebooks expect model checkpoints to be attached as Kaggle datasets:

• tokiggle/clara-7b-e2e-4q (Apple E2E pretrained checkpoint)
• tokiggle/clara-ft-squad and tokiggle/clara-ft-triviaqa (fine-tuned checkpoints uploaded from this project)

Code map (quick)

• models/ — reimplementation of CLaRa core and helpers.
• data/ — unified dataset loader.
• configs/ — training config defaults.
• scripts/ — Stage I/II training and Apple-native fine-tune/eval drivers.
• modeling_clara.py — patched Apple model code used by the Kaggle pipelines. The original lives inside the Apple checkpoint, which is not tracked in this repo's git.