Arushishukla/anlp-a2-part2-optimizers

ANLP A2 Part 2 -- Optimizers

Checkpoints for ANLP Assignment 2 (Monsoon 2026). One subfolder per run; each contains best.pt (state dict + config, selected by lowest validation loss) and its evaluation files. tokenizer.json at the repo root is shared by every run.

run	n_params	test perplexity	test BLEU	tokens seen
adamw	16,765,824	88.484	0.50	39,373,632
cautious	16,765,824	93.595	0.76	39,373,632
lion	16,765,824	97.194	0.63	39,373,632
muon	16,765,824	66.498	0.71	39,373,632

Load a checkpoint:

import torch
ckpt = torch.load('best.pt', map_location='cpu', weights_only=False)
# ckpt['model_state'], ckpt['config'], ckpt['step'], ckpt['tokens_seen']