================================================================================
hyperftype.json — train / validation / test split (documentation)
================================================================================

WHAT THIS FILE IS
  The experiments described in this portfolio (Temporal Separability Benchmark,
  downstream GRU training, and comparisons to the single-frame baseline) use a
  single, frozen partition of the 3rd Aptos training examinations into train,
  validation, and test subsets. That partition is stored as a JSON file
  conventionally named:

    hyperftype.json

WHERE THE REAL FILE LIVES
  The authoritative copy is kept on the Idiap research infrastructure together
  with the benchmark and training code paths used during the internship — not in
  this public graduation folder. Reasons:

  • The split is tied to internal experiment scripts and directory layouts on the
    cluster.
  • Examination identifiers in the split refer to dataset entries whose full
    redistribution is governed by the Aptos competition terms and local policy;
    shipping the exact ID list in a student portfolio is unnecessary for assessment
    and was avoided unless explicitly required.

  Supervisors and internal collaborators can reproduce runs by using the same
  hyperftype.json from the project storage location configured in the training
  repository.

WHAT THE JSON REPRESENTS (LOGICAL STRUCTURE)
  The file groups examination identifiers (or folder names, depending on the
  loader version) into three disjoint sets. Typical top-level shape:

  {
    "train":   [ ... list of exam ids / folder names ... ],
    "val":     [ ... ],
    "test":    [ ... ]
  }

  Alternative equivalent layouts sometimes use keys such as "validation" instead
  of "val"; the loader in the Idiap codebase maps these to the same three-way
  split. All models compared in the thesis work were evaluated on the same
  held-out test set (211 examinations in the configuration documented in the
  design decision record).

REDACTED EXAMPLE (NOT REAL DATA — FOR SCHEMA ILLUSTRATION ONLY)
  The following is a fabricated miniature example. Do not use these IDs; they
  are placeholders to show structure only.

--------------------------------------------------------------------------------
{
  "train": ["EXAM_PLACEHOLDER_0001", "EXAM_PLACEHOLDER_0002", "EXAM_PLACEHOLDER_0003"],
  "val":   ["EXAM_PLACEHOLDER_0004", "EXAM_PLACEHOLDER_0005"],
  "test":  ["EXAM_PLACEHOLDER_0006", "EXAM_PLACEHOLDER_0007"]
}
--------------------------------------------------------------------------------

HOW THIS PORTFOLIO REFERENCES IT
  main.tex and 04_Design/design-decisions.tex cite hyperftype.json to state that
  metrics are comparable across configurations because the split is fixed. For
  the graduation hand-in, this README is the canonical note on where the file
  lives and what it contains. If an assessor needs verification, the student can
  provide the path on Idiap or a supervisor confirmation without publishing raw
  identifiers here.

A machine-readable redacted example with the same keys is also stored as:

    06_Appendices/hyperftype-split-sample.json

Last updated: April 2026 — Tymo van Rijn