My attempt of the summer '26 assignment.
Chosen track (other than mandatory Core ML): Mechanistic Interpretability.
Refer to Report.pdf for analytical details; code is explained through in-line comments.
.
├── Core ML
│ ├── Data
│ │ ├── Attention Convolution Hybrids/
│ │ ├── Attention Variants/
│ │ └── Positional Embeddings/
│ └── Notebooks
│ ├── attention_variants.ipynb
│ ├── baseline.ipynb
│ ├── convolution_attention_hybrids.ipynb
│ └── positional_embeddings.ipynb
├── Mechanistic Interpretability
│ ├── Data
│ │ ├── m1024/
│ │ ├── m512/
│ │ └── normalizer.pt
│ ├── Notebook Logs
│ │ ├── m1024/
│ │ ├── m512/
│ │ └── pipeline-setup.log
│ └── Notebooks
│ ├── m1024/
│ ├── m512
│ │ ├── damage-robust-quantisation.ipynb
│ │ ├── quantisation-analysis.ipynb
│ │ └── sae-training-m512.ipynb
│ └── pipeline-setup.ipynb
└── Report.pdf