Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
1.04k
2.09k

Patch-Aliasing Bayesian Analysis Data

Raw and processed data from the Bayesian analysis of structural patch-aliasing in Chronos-Bolt Tiny. Companion to the model weights at federicosabbadini/chronos-bolt-patch-aliasing-models.

Structure

clean_15model/          # 15 preregistered (P,S) configurations
full_22model/           # all 22 configurations (adds 7 robustness checks)
h1_fixed_offset/        # supplementary H1 analysis with fixed offset

Each run folder contains:

  • 01_prior_spec.json / .parquet — prior specification
  • 03_recovery.parquet — parameter recovery test
  • 04_*.nc — MCMC trace files (NetCDF/ArviZ InferenceData)
  • 05_diagnostics.parquet — convergence diagnostics
  • 05_sensitivity.parquet — prior sensitivity analysis
  • 05_verdicts.parquet — hypothesis verdicts
  • data/ — input data (contrasts, collapse profiles, MDL cells, sites)
  • data/signals/ — generated probe signals
  • figures/ — analysis figures (prior predictive, forests, combs, PPC, sensitivity)
  • run.log — execution log

Models fitted

ID Estimand Hypothesis
A Population forecast attenuation H1 behavioural
B Codelength expansion at locked frequencies H1 representational
C Phase-offset dispersion H2
D1 Token-collapse grid labelling H3a (stride branch)
D2 Stride-scaling slope H3a (κ_S)

Key results

  • H1 behavioural: inconclusive (forecast contrast 1.08 [0.45, 2.49])
  • H1 representational: supported on clean (codelength 1.27 [1.23, 1.31])
  • H2 phase invariance: supported (σ_φ = 0.033, ROPE prob 0.997)
  • H3a stride branch: supported (κ_S = 1.000 [0.992, 1.008])
  • H3b patch branch: not identified

Trace file sizes

The .nc files contain full MCMC traces (4 chains × 2000 draws). Models A and C are large (~8 GB each) due to per-observation random effects; Models B, D1, D2 are smaller (4–730 MB).

License

Apache-2.0 — same as the Chronos models these data derive from.

Downloads last month
46