Dispersion labels form a distribution: calibrated widths score annotation reliability

Public reproducibility release for the IEEE TGRS submission prepared on 19 November 2026. The release contains the frozen waveform-conditioned Flow, calibration, five independently trained ensemble members, all-waveform training partitions, point-level evaluation arrays, conventional two-step inversion inputs and outputs, numerical evidence, and figure-generation sources.

All assets are downloadable without an account or access approval. The dataset remains pinned at SeisDispFusion-NCF revision 1c93243311b39bd255b7c9d4072fa4779cda3045, DOI 10.57967/hf/9114. Its waveform-only and dispersion-image HDF5 files remain in that public dataset repository; they are not unnecessarily duplicated here.

Frozen results and scope

All 51,409 training waveforms from the 74,114-waveform archive participate, including unlabelled waveforms through reconstruction. The model describes labels conditional on waveforms across 49 periods. Exact marginal 90% intervals attain 91.36% held-out coverage, and the fixed 5% relative-width rule retains 333,808 test candidates in the 5–40 s domain. Screening uses marginal reliability; the observed joint contract is separately reported as 75.87–80.39%.

The aligned-neighborhood extension contains 128 centers, 53 shared-path groups, and 3,335 period points in its principal arm. All six aligned definitions have positive period-adjusted width–scatter point estimates, spanning 0.098–0.312. The earlier strict zero-lag single-center diagnostic remains an explicitly identified sensitivity baseline.

Flow width attains absolute-residual AUC 0.700 [0.655, 0.742]. The separate relative-residual event uses an identical frozen event for all methods: Flow AUC 0.698 [0.656, 0.742], ensemble AUC 0.663 [0.622, 0.707], and paired gain 0.035 [0.018, 0.052]. Both inversion arms use identical original label values, solvers and regularization; only the posterior-selection mask changes. v2.2 is a compatibility reference.

Download and verify

Use the immutable scientific release tag tgrs-20261119 (the exact commit is recorded in the submission's Data and code availability statement). Run downloads on your compute host:

from huggingface_hub import snapshot_download
snapshot_download("cangyeone/ncf-annotation-distributions",
                  revision="tgrs-20261119", local_dir="ncf-release",
                  token=False)

The original pre-extension release remains available at commit ff305f57ffdddb10ec0d1ed1f68bc2ee6657458e; its root-level SHA256SUMS.json describes that historical release only. The current scientific release is indexed by TGRS_SHA256SUMS.json, PACKAGE_SHA256.json, and RUN_FILES_SHA256.json.

Create a Python environment matching environment.json and requirements-analysis.txt. The tested scientific environment is Python 3.13 and the recorded PyTorch CUDA build. Install the matching PyTorch wheel from its CUDA index when needed; evaluation from archived arrays runs on CPU. CUDA is required by the original training/inference scripts.

python ncf-release/unpack_release.py --release ncf-release --workspace ncf-work
python ncf-release/verify_release.py --runs ncf-work/runs --output ncf-check

The verifier checks split membership counts, inclusion of unlabelled training waveforms, model/calibration hashes, marginal coverage, width, retention, ensemble calibration, and the scan maximum. It also reruns the original paired cluster bootstrap on the archived arrays and requires exact equality with the complete E1/E2 AUC JSON. reproduction_check/verification.json records the successfully executed check on CUDA's host CPU.

Reproduce analyses and figures

run_archived.py relocates the historical /mnt/e/NCFDispFlow path in memory, leaving source files and their recorded hashes unchanged. Always use an isolated extracted workspace: original programs write their output files. For example:

python ncf-release/run_archived.py --runs ncf-work/runs \
  --entry 20261119_tgrs_v1/code/tgrs_evaluate.py
python ncf-release/run_archived.py --runs ncf-work/runs \
  --entry 20261119_tgrs_v1/code/plot_tgrs_figures.py --font 'DejaVu Sans'

The published figure PDFs are the submission versions. Arial font binaries are not redistributed. Install a licensed Arial font for identical typography, or use the explicit fallback above; plotting quantities do not depend on the font. Figure 2 is an Inkscape SVG composition; its editable submission source is included with paper_assets/figures/. Restoring its three corner labels was a presentation-only edit.

For full inference/analysis reruns and retraining, use the archived training, calibration, exact-quantile, neighborhood, intervention, balancing and conventional inversion scripts listed in REPRODUCIBILITY_MAP.md. Preserve the frozen partitions and calibration/test separation. The raw-image-dependent scripts expect the two dataset HDF5 files under runs/20261115_pilot/data/; download them from the pinned dataset into that location on the compute host. Prepared banks and frozen predictions are already supplied for the primary reproduction path.

Package layout

  • packages/20261115_unfiltered_v2.tar.gz: complete prepared bank, split assignments, original checkpoint/calibration, exact posterior arrays, and historical sampling comparison.
  • packages/20261115_qc_information_v1.tar.gz: waveform interventions and exact-QC diagnostics.
  • packages/20261115_balance_v1.tar.gz: matched-compute three-arm checkpoints, calibration, predictions and summaries.
  • packages/20261117_traditional_label_qc_v1.tar.gz: original-value two-step inversion inputs, sparse operator, phase/depth fields and validation.
  • packages/20261119_narrative_v1.tar.gz, packages/20261119_neighbor_diagnostic_v1.tar.gz, packages/20261119_three_leg_v1.tar.gz: analysis arrays, neighborhood membership/alignment, protocols, bootstrap outputs and scans.
  • packages/20261119_tgrs_v1.tar.gz: all five ensemble checkpoints/configurations/training histories, mixture calibration/predictions and paired evaluation.
  • Other dated packages: extraction examples and figure inputs; supporting_inputs contains upstream preparation dependencies and the frozen v2.2 comparison, not superseded probabilistic inversion experiments.
  • paper_assets.tar.gz: submission figure assets and machine-readable manuscript evidence.
  • publication/ (on the repository's current main revision): the finalized manuscript, SI and scientific LaTeX sources. These document files cite the pinned scientific release. Cover letters and internal author checklists are not part of the public scientific archive.

Attribution and use

Retain the dataset attribution and any third-party notices. This publication makes the research artifacts publicly accessible; it does not relicense third-party datasets, fonts or libraries. See the dataset's own card for its applicable terms.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train cangyeone/ncf-annotation-distributions