ad1tya2/anlp_assignment_1

ANLP Assignment 1 - Transformer ablation (C1-C5)

Encoder-decoder Transformers built from basic PyTorch operations, mapping encrypted binary sequences to plaintext English. Five configurations differing in exactly one component each: positional encoding, attention, normalization, and tokenization (C5 is a simplified Byte Latent Transformer).

Config	Params	Bit acc	Char acc	Seq acc	Levenshtein	BLEU
C1	6,581,248	0.9485	0.9979	0.628	1.02	98.49
C2	6,581,248	0.9718	0.9978	0.632	1.24	98.50
C3	5,693,056	0.9356	0.9968	0.524	1.60	97.75
C4	6,576,896	0.9453	0.9979	0.632	0.99	98.49
C5	6,391,427	0.9933	0.9851	0.146	7.59	--

Checkpoints are raw state_dict files. Rebuild the model with src/config.py and src/models/, then load_state_dict. Tokenizers are under tokenizers/.