PhyGHT mu60
Checkpoint of PhyGHT. Training completed all 200 epochs on 8,000 events. The checkpoint was selected at epoch 192 using validation hybrid loss only. Test results were measured after training and selection. These measurements are not paper results or a claim of reproduction.
| Split | Target | R2 | MAE | RMSE | MSE |
|---|---|---|---|---|---|
| validation | energy_fraction | 0.93763456 | 0.02846976 | 0.04437589 | 0.00196922 |
| validation | mass_fraction | 0.86227840 | 0.02567715 | 0.03930365 | 0.00154478 |
| test | energy_fraction | 0.94451230 | 0.02774572 | 0.04326318 | 0.00187170 |
| test | mass_fraction | 0.86913643 | 0.02548577 | 0.03961160 | 0.00156908 |
Full results and provenance are in metrics.json, config.json, and
source_provenance.json. Complete training history, the training log, and the
environment record are included.
Install and infer
The verified runtime is Python 3.10.18 with CUDA-enabled PyTorch 2.6.0+cu124. Install locally in an isolated environment:
python -m pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
python -m pip install torch-geometric==2.6.1 numpy==1.26.4 safetensors huggingface-hub
python -m pip install --only-binary=:all: 'torch-cluster==1.6.3+pt26cu124' 'torch-scatter==2.1.2+pt26cu124' -f https://data.pyg.org/whl/torch-2.6.0+cu124.html
import json, sys, os
os.environ.setdefault("CUBLAS_WORKSPACE_CONFIG", ":4096:8")
from pathlib import Path
import torch
from huggingface_hub import snapshot_download
folder = Path(snapshot_download("Rakib/PhyGHT-mu60"))
sys.path.insert(0, str(folder))
from portable_io import load_model, predict_event
device = "cuda" if torch.cuda.is_available() else "cpu"
model = load_model(folder, device=device)
event = json.loads((folder / "example_event.json").read_text())
predictions = predict_event(model, event, device=device) # [number of jets, 2]
Inputs and preprocessing
One event has jets (float64 rows [pT, eta, phi, mass, Efrac, Mfrac]), trks
(float64 rows [pT, eta, phi, mass, d0, z0, track_id, truth_label]), and
jet_trk_IDs (variable-length lists of int64 identifiers, one per jet). Model
features use only the first six track and first four jet columns, cast to
float32 without normalization or truncation. Track ID column 6 identifies
constituents; identifiers are not track positions. Truth labels and regression
targets are excluded from features. At inference, jet rows may contain only
four feature columns and track rows seven columns, including track ID.
The auxiliary signal target is truth_label == -1. Training uses energy MSE +
mass MSE + 0.1 times track BCE. The model has dimension 128, four heads, three
Transformer layers, eight neighbors, dropout 0.1, and AdamW with learning rate
0.0003, weight decay 0.01, and gradient clipping 1.0. Effective batch size is 16.
Smaller memory microbatches, if needed, retain per-jet MSE and per-track BCE
weighting before one clipped optimizer update per logical batch.
Attribution, license, and limitations
Authors: Mohammed Rakib, Luke Vaughan, Shivang Patel, Flera Rizatdinova, Alexander Khanov, and Atriya Sen.
- Paper: https://arxiv.org/abs/2602.20475
- Code: https://github.com/rAIson-Lab/PhyGHT at
ce3a02873bbe8b6382f7ba878d3866ff2aecb203 - Original simulation data: https://doi.org/10.5281/zenodo.18746388
- Dataset redistribution: https://huggingface.co/datasets/Rakib/PhyGHT-TopQuark
- Downloads last month
- 48