anlp-a2-MoE-var3

Advanced NLP (Monsoon 2026), Assignment 2.

Part 1, FFN variant 3. 8.8M total / 7.3M active params, trained on 120000840 tokens. Best checkpoint by validation loss; test loss 2.057, ppl 7.821, acc 0.625, bleu 22.250.

Run

item	value
parameters	8.8M
training step	89078
wandb	https://wandb.ai/winterdew-iiit-hyderabad/anlp-a2-part1/runs/kxvhww45

Hyperparameters

setting	value
FFN variant	3 — 4 experts, 2 active per token, total-param matched
params (total / active)	8.8M / 7.3M
FFN hidden width	256 per expert
MoE aux weight	0.01
d_model / n_layers / n_heads	256 / 6 / 8
context length	256
positional encoding	RoPE (base 10000)
tokenizer	sentencepiece unigram, vocab 16000, character coverage 0.9995, byte fallback
optimizer	AdamW(betas=(0.9, 0.95), eps=1e-8)
learning rate	0.0003
lr schedule	linear warmup 1740 steps, cosine to 0.1x
weight decay	0.1 (matrices only)
gradient clip	1.0
batch size	32 x 1 accum
dropout	0.1
precision	fp16
token budget	120,000,000 (120,000,840 seen)
seed	26

Model config

{
  "vocab_size": 16000,
  "n_ctx": 256,
  "rope_base": 10000.0,
  "d_model": 256,
  "n_heads": 8,
  "n_layers": 6,
  "d_ff": 1024,
  "dropout": 0.1,
  "tie_embeddings": true,
  "use_moe": true,
  "n_experts": 4,
  "n_active": 2,
  "n_shared": 0,
  "match_active": false,
  "moe_aux_weight": 0.01
}