ANLP Assignment 1: Transformers from Scratch & Ablation Study
This repository contains trained model checkpoints and evaluation outputs for ANLP Assignment 1 across five controlled architectural configurations:
- C1 (Base): Sinusoidal Absolute PE + Multi-Head Attention (MHA) + Pre-LayerNorm + Scratch BPE
- C2 (RoPE): Rotary Positional Embedding (RoPE) + MHA + Pre-LayerNorm + Scratch BPE
- C3 (GQA): Sinusoidal Absolute PE + Grouped-Query Attention (GQA) + Pre-LayerNorm + Scratch BPE
- C4 (RMSNorm): Sinusoidal Absolute PE + MHA + RMSNorm + Scratch BPE
- C5 (BLT): Sinusoidal Absolute PE + MHA + Pre-LayerNorm + Token-Free Byte Latent Transformer
Experimental Results Summary
----------------------------------------------------------------------------------------------------------------
ANLP ASSIGNMENT 1 β ABLATION STUDY RESULTS
----------------------------------------------------------------------------------------------------------------
Config | Status | Val Loss | Seq Acc | Bit Acc | Levenshtein | BLEU | VRAM (GB) | Time (s)
----------------------------------------------------------------------------------------------------------------
C1 | SUCCESS | 3.2377 | 0.0% | 36.8% | 843.3 | 0.03 | 5.30 | 1043.4
C2 | SUCCESS | 2.8871 | 0.0% | 40.7% | 641.0 | 0.13 | 5.36 | 1081.6
C3 | SUCCESS | 3.2527 | 0.0% | 36.8% | 844.5 | 0.01 | 5.35 | 999.9
C4 | SUCCESS | 3.2459 | 0.0% | 36.9% | 839.1 | 0.04 | 5.19 | 1009.3
C5 | SUCCESS | 0.0024 | 0.0% | 0.0% | 577.5 | N/A | 0.87 | 182.4
----------------------------------------------------------------------------------------------------------------
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support