Aryama/anlp-a2-m26

ANLP Assignment 2 checkpoint bundles

Each checkpoint subfolder contains model.pt, config.json, tokenizer.json and metadata.json. Each metadata.json records the SHA-256 of its weights. The custom architecture implementation is included under src/. These are PyTorch bundles.

Download this repository, install its locked environment with uv sync --frozen, and load a checkpoint from its root directory:

from src.common.checkpoint import load_bundle
model, config, metadata = load_bundle("LLM-experiments-no_45/final", device="cpu")

Final checkpoints used in the report:

• Part 1 (final translation weights, peak learning rate 3e-4): 45 dense, 46 top-1 of 4, 47 top-2 of 4, 48 shared + top-1 of 3, 49 active-matched top-2.
• Part 2 (milestone-01 to milestone-10 at 0.1-corpus intervals): 77 AdamW, 78 NAdamW, 79 Muon, and 66 Lion with tuned settings (lr 1e-4, weight decay 1.0, betas 0.95/0.98).

All checkpoints were trained on Modal L40S GPUs.

Part 3 uses the existing pretrained Pythia-160M checkpoint and adds no trained model. This file provides loading instructions, not a writeup.