MLPF as a foundation model — downstream evaluation dataset (FCC-ee CLD, e⁺e⁻ → tt̄, √s = 365 GeV)

This dataset accompanies the paper "Machine-learned particle flow as a foundation model for collider physics" (arXiv:2606.14373). It contains the preprocessed test and validation sets used to produce all figures and tables in that paper. The companion code is available at https://github.com/farakiko/mlpf-foundation. --- PHYSICS CONTEXT --- Events are simulated e⁺e⁻ → tt̄ collisions at √s = 365 GeV in the CLD detector geometry at FCC-ee. Reconstruction is performed by MLPF, a transformer-based particle-flow model fine-tuned on the CLD Key4hep simulation. The MLPF backbone is held frozen throughout all downstream experiments. All events in this dataset are drawn exclusively from the 4,218 file held-out test split in Ref. Phys. Rev. D 111, 092015. --- WHAT IS IN THIS RECORD --- The record contains four data folders at two pipeline stages, plus one PCA model artifact: data/enriched_paper_test/downstream_test (~2.0 GB, 400 files) data/enriched_paper_test/downstream_val (~255 MB, ~50 files) data/jet_processed_paper_test/downstream_test (~1.5 GB, ~400 files) data/jet_processed_paper_test/downstream_val (~189 MB, ~50 files) preprocessing/pca_cld.npz (~104 MB) Each data file is a Parquet file containing ~100 events, readable with pandas or awkward-array. All arrays are jagged (variable length per event) unless noted. --- STAGE 1: ENRICHED PARQUETS (event-level) --- Produced by preprocessing/postprocess_root.py followed by preprocessing/extract_latents.py. This format is used directly for MET regression, and subsequently passed through preprocessing/build_jet_dataset.py to produce the jet-processed parquets used for b-tagging and jet energy regression. Column dtype shape per event description ------ ----- --------------- ----------- pt float32 (N_pfcands,) MLPF-reconstructed pfcand pT [GeV] eta float32 (N_pfcands,) pfcand pseudorapidity sin_phi float32 (N_pfcands,) sin of pfcand azimuthal angle cos_phi float32 (N_pfcands,) cos of pfcand azimuthal angle energy float32 (N_pfcands,) pfcand energy [GeV] pid int32 (N_pfcands,) MLPF reconstructed PID class: 1=chHad, 2=neutHad, 3=photon, 4=electron, 5=muon pid_probs float32 (N_pfcands, 5) soft PID probabilities over the 5 classes above (row sums to 1) elem_type int32 (N_pfcands,) MLPF input element type (track vs. calorimeter cluster) latent_cld float32 (N_pfcands, 128) 128-dim PCA projection of the frozen MLPF per-particle latent reps track_params float32 (N_pfcands, 5) helix parameters at the inner tracker point (AtIP) [D0, phi0, omega, Z0, tan(lambda)] track_cov float32 (N_pfcands, 15) 5×5 helix covariance matrix track_d0_pv float32 (N_pfcands,) D0 corrected to the primary vertex via full Lorentz-force helix [mm] track_z0_pv float32 (N_pfcands,) Z0 corrected to PV [mm] pv_x/pv_y/pv_z float64 (1,) smeared primary vertex position [mm] genmet float64 (1,) generator-level MET magnitude [GeV] genmet_x float64 (1,) generator-level MET x-component genmet_y float64 (1,) generator-level MET y-component genmet_z float64 (1,) generator-level MET z-component genjets float64 (N_genjets, 4) generator jets: [pt, eta, phi, E] targetjets float64 (N_targetjets, 4) target jets for JER (matched reco jets): [pt, eta, phi, E] bhadron_pt/eta/ float32 (N_B,) last-copy B-hadron kinematics [GeV] phi/energy dhadron_pt/eta/ float32 (N_D,) last-copy D-hadron kinematics [GeV] phi/energy --- STAGE 2: JET-PROCESSED PARQUETS (jet-level) --- Produced by preprocessing/build_jet_dataset.py, which clusters pfcands with the ee-anti-kT algorithm (R=0.4, fastjet), matches to gen-jets (ΔR < 0.4), and assigns flavor labels via ghost association. Each file contains ~575 jets. All arrays are jagged (variable length per jet, i.e. variable number of constituents) unless noted as scalar. Column dtype shape per jet description ------ ----- ------------- ----------- feats float32 (N_const, 36) per-constituent feature matrix aligned with (arXiv:2202.03285, Table 1). Feature index layout: 0 thetarel polar angle wrt jet axis in jet rest frame [rad] 1 phirel azimuthal angle in jet rest frame 2 log_pt log1p(pt) 3 log_E log1p(energy) 4 log_pt_rel log(pt/jet_pt) 5 log_E_rel log(E/jet_E) 6 dR lab-frame ΔR to jet axis 7 d0 helix d0 wrt PV (=track_d0_pv) [mm] 8 z0 helix z0 wrt PV (=track_z0_pv) [mm] 9 sip2d jet-signed 2D impact parameter 10 sip2d_sig sip2d/σ(d0); 11 sip3d jet-signed 3D IP 12 sip3d_sig sip3d/σ(3D); sentinels as above 13 d3d signed skew-line distance to jet axis through PV 14 d3d_sig d3d/σ; sentinels as above 15–29 track covariance C_ij 30 charge sign(omega) for tracks; 0 for clusters 31 pid_chHad one-hot pid==1 32 pid_neutHad 33 pid_photon 34 pid_e 35 pid_mu latent_cld float32 (N_const, 128) per-constituent MLPF latent (same 128-dim PCA projection as in Stage 1, subset to jet constituents) label int8 scalar flavor label: 0 = light (udsg), 1 = c-jet, 2 = b-jet assigned by ghost association with b > c > light priority jet_pt float32 scalar reco jet pT [GeV] (fastjet) jet_eta float32 scalar reco jet η jet_phi float32 scalar reco jet φ, wrapped to (−π, π] jet_E float32 scalar reco jet energy [GeV] gen_jet_pt float32 scalar matched gen-jet pT [GeV] gen_jet_eta float32 scalar matched gen-jet η gen_jet_phi float32 scalar matched gen-jet φ gen_jet_E float32 scalar matched gen-jet energy [GeV] n_const int32 scalar number of real constituents (= first dimension of feats) --- PCA MODEL (preprocessing/pca_cld.npz) --- A scikit-learn IncrementalPCA model serialised with numpy.savez_compressed. It was fit on the pca_fit slice (separate from test/val) using the 2048-dimensional raw MLPF latent vectors, retaining the 128 components used as input features for all downstream tasks.