Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: pytorch
|
| 3 |
+
tags: [transformer, ablation, anlp]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# ANLP Assignment 1 — Architectural Ablation
|
| 7 |
+
|
| 8 |
+
Encoder-decoder transformers built from scratch for a binary-cipher →
|
| 9 |
+
plaintext sequence task. Five configurations differing by exactly one
|
| 10 |
+
component each.
|
| 11 |
+
|
| 12 |
+
| Config | Positional | Attention | Norm | Tokenization |
|
| 13 |
+
|---|---|---|---|---|
|
| 14 |
+
| C1 | Sinusoidal | MHA | LayerNorm | Subword BPE |
|
| 15 |
+
| C2 | **RoPE** | MHA | LayerNorm | Subword BPE |
|
| 16 |
+
| C3 | Sinusoidal | **GQA** | LayerNorm | Subword BPE |
|
| 17 |
+
| C4 | Sinusoidal | MHA | **RMSNorm** | Subword BPE |
|
| 18 |
+
| C5 | Sinusoidal | MHA | LayerNorm | **BLT (token-free)** |
|
| 19 |
+
|
| 20 |
+
Uploaded: C1, C2, C3, C4, C5
|
| 21 |
+
|
| 22 |
+
Trained on IIIT-H Ada (RTX 2080 Ti, 11 GB).
|