anuml/anlp-a3-part1-cipher

ANLP Assignment 3 Models: PART1

Model checkpoints and assets for ANLP Assignment 3 (PART1).

Checkpoint Architecture

• Formulation: Decoder-Only Prefix-LM mapping cipher inputs to plain characters.
• Architecture: Decoder-Only Transformer (RMSNorm, RoPE, GQA) & Mamba SSM.
• Tokenization: Character-level tokenizer with fixed 86-token alphabet.

Uploaded Assets

Validation Benchmarks

Model	Epoch	Val Loss	Char Acc	Word Acc	F1	Levenshtein	Params
transformer	3	1.1532	0.0690	0.0479	0.4203	780.01	2,916,096

Synchronization & Usage

Download these checkpoints locally using the provided script:

python3 scripts/sync_models.py download --part part1

Or using the shell script:

bash scripts/sync_models.sh download --part part1

Python Loading Example

import torch
from src.models.transformer import DecoderOnlyTransformer

ckpt = torch.load('models/part1/transformer/best_model.pt', map_location='cpu')
config = ckpt['config']
model = DecoderOnlyTransformer(**{k: v for k, v in config.items() if k not in ('model_type', 'gradient_checkpointing')})
model.load_state_dict(ckpt['model_state_dict'])
model.eval()