dynamic_sizing fields

[jviquerat]

Hi,

First, thanks for this repo.
I may be wrong, but for datasets flag_dynamic_sizing and sphere_dynamic_sizing (download through the .sh script), the dynamic_sizing field in the .json file has strange attributes:

"sizing_field": {
      "type": "dynamic_varlen",
      "shape": [
        -1,
        4
      ],
      "dtype": "<dtype: 'float32'>"
    }

This raises a first attribute error due to the malformed dtype, then a second reshape error.
Modifying to:

"sizing_field": {
      "type": "dynamic_varlen",
      "shape": [
        -1,
        3
      ],
      "dtype": "float32"
    }

seems to solve the problem.

[NamanAnand66]

Fix malformed dtype and incorrect shape in dynamic sizing fields (#651)

🧩 Summary

This PR fixes malformed dtype and incorrect shape fields found in the metadata JSON for the following datasets:

• flag_dynamic_sizing
• sphere_dynamic_sizing

---

🐞 Problem

The affected JSON metadata currently includes:

"sizing_field": {
  "type": "dynamic_varlen",
  "shape": [
    -1,
    4
  ],
  "dtype": "<dtype: 'float32'>"
}

When attempting to load the affected datasets, users encounter two main issues:

1. Attribute Error:
   TensorFlow / NumPy cannot interpret the malformed dtype string
   <dtype: 'float32'> as a valid data type.
2. Reshape Error:
   The tensors contain only 3 elements per sample, but the metadata specifies a shape of [ -1, 4 ].
   This mismatch leads to reshape failures during loading.

---

✅ Fix

The sizing_field definition was updated to correctly reflect the data format:

"sizing_field": {
  "type": "dynamic_varlen",
  "shape": [
    -1,
    3
  ],
  "dtype": "float32"
}

Changes:

• Fixed dtype from "<dtype: 'float32'>" to "float32"
• Corrected shape from [-1, 4] to [-1, 3] to match actual tensor dimensions