ANLP Assignment 1: Transformer Ablations
This repository contains the trained weights for a sequence-to-sequence translation task using various custom-built Transformer architectures from scratch.
Checkpoints Included
C1_baseline/best.pt: Baseline Transformer (Sinusoidal PE, MHA, LayerNorm)C2_rope/best.pt: Rotary Positional Encoding (RoPE)C3_gqa/best.pt: Grouped-Query Attention (GQA)C4_rmsnorm/best.pt: RMSNorm
Tokenizers
The shared BPE tokenizers used to encode the sequence data are included in the root directory:
src_tokenizer.jsontgt_tokenizer.json
Author: Rachit Mehta (2024101089)