anlp-a2-optimizers-adamw

Advanced NLP (Monsoon 2026), Assignment 2.

Part 2, optimizer adamw. 10.4M params, trained on 42090496 tokens = 0.33x Chinchilla. Best checkpoint by validation loss; test loss 4.342, ppl 76.845, acc 0.280, bleu 0.486.

Run

item	value
parameters	10.4M
training step	5138
wandb	https://wandb.ai/winterdew-iiit-hyderabad/anlp-a2-part2/runs/jw32g7nc

Hyperparameters

setting	value
optimizer	AdamW
max_grad_norm	1.0
weight_decay	0.1, 0.0
betas	(0.9, 0.95)
eps	1e-08
lr	0.0003
param_groups	2
trainable_params	10391808
schedule	linear warmup 2% then cosine to 10% of peak
batch_size	32
grad_accum	1
tokens_per_step	8192
n_ctx	256
total_steps	5137
warmup_steps	102
trained_tokens	42090496
chinchilla_ratio	0.3342
precision	fp16
seed	26

Model config

{
  "vocab_size": 16000,
  "n_ctx": 256,
  "rope_base": 10000.0,
  "d_model": 256,
  "n_heads": 8,
  "n_layers": 8,
  "d_ff": 1024,
  "dropout": 0.1,
  "tie_embeddings": true,
  "use_moe": false,
  "n_experts": 4,
  "n_active": 1,
  "n_shared": 0,
  "match_active": false,
  "moe_aux_weight": 0.01
}