SeismoGPT (Phase 1)

Pre-trained Phase 1 weights for SeismoGPT: a causal RoPE transformer over seismic waveform tokens (Z, N, E), trained with log-cosh loss and multi-horizon auxiliary heads.

This checkpoint is the deterministic model used for paper rollouts and figures in the companion code repository.

Paper: Data-Driven Forecasting of three-Component Seismograms Using Transformer Architectures — arXiv:2606.02912v1
Code: github.com/wesmail/SeismoGPT

Files

File	Description
epoch=12-step=633750.ckpt	PyTorch Lightning checkpoint (Phase 1, epoch 12)
train_phase1_logcosh.yaml	Training configuration used for Phase 1

Requirements

• Python 3.10+
• torch, lightning, seisbench, and the SeismoGPT code repository

Download weights

pip install huggingface_hub
huggingface-cli download wesmail/SeismoGPT \
  epoch=12-step=633750.ckpt train_phase1_logcosh.yaml \
  --local-dir phase1

Or from Python:

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download(
    repo_id="wesmail/SeismoGPT",
    filename="epoch=12-step=633750.ckpt",
    local_dir="phase1",
)

Use the same kernel_size (16) and num_tokens (320) as in training; these are stored in the checkpoint hyperparameters. For evaluation (configurations A/B/C), rollouts, and plotting, see the code repository.

Model summary

Setting	Value
Phase	1 (deterministic, log_cosh)
Channels	Z, N, E
Token size K	16 samples
Tokens per window	320
d_model	512
Encoder layers	8
Prediction horizons (train)	4 (inference uses horizon 1)