KyleBae1017/stormer-decay-checkpoints

Stormer — Decay Finetuning Checkpoints

PyTorch Lightning checkpoints from a Stormer decay-finetuning sweep (h512_d20_n8: hidden size 512, depth 20, 8 heads).

Each subfolder corresponds to one run and contains:

• checkpoints/epoch_*.ckpt — best checkpoint (monitored on val/w_mse_aggregate_72_hrs_ensemble_mean)
• checkpoints/last.ckpt — last checkpoint (identical to the epoch checkpoint for every run except decay_h512_d20_n8_ep9_d1)
• config.yaml — the full LightningCLI training config

Runs

Folder	Epochs	decay (d)
decay_h512_d20_n8_ep9_d1	9	1
decay_h512_d20_n8_ep18_d2	18	2
decay_h512_d20_n8_ep36_d4	36	4
decay_h512_d20_n8_ep54_d6	54	6
decay_h512_d20_n8_ep72_d8	72	8
decay_h512_d20_n8_ep90_d10	90	10
decay_h512_d20_n8_ep108_d12	108	12

Loading

import torch
from huggingface_hub import hf_hub_download

ckpt = hf_hub_download(
    "KyleBae1017/stormer-decay-checkpoints",
    "decay_h512_d20_n8_ep108_d12/checkpoints/last.ckpt",
)
state = torch.load(ckpt, map_location="cpu")