Fix corrupt seed-2 shuffle index

#2
by stellaathena - opened

The published seed-2 training shuffle_idx.npy contains a large zero-filled region and cannot reproduce the intended PolyPythia data order.

Problem

  • Existing LFS SHA-256: 70636bdf9f8ac1343a1aa6c6b0da7eab2983b80855f1b452d99e158c7519ab21
  • Shape/dtype: (162165686,), uint32
  • Corrupt interval: [12354528, 162164736)
  • Corrupt length: 149,810,208 consecutive zeros (92.38% of the array)

Because the supplied loader applies idx = shuffle_idx[idx], every entry in this interval resolves to packed sample 0. Across the 146,432,000 samples consumed by a 143,000-step run, this would repeat sample 0 for 134,077,472 positions (91.56%).

Fix

Replace only the corrupt file with the intact deterministic seed-2 permutation.

  • New SHA-256: 1e0844e29f86cb1113fbb63d8f304ee1a41734ba89d915c6957eee76b20dcdc9
  • Shape/dtype: (162165686,), uint32
  • Minimum/maximum: 0 / 162165685
  • Number of zero entries: 1
  • Sum: 13148854776842455, equal to the sum of [0, 162165686)

I independently regenerated the expected maps with np.random.RandomState(seed=2): first shuffling the 210,604,984-entry doc_idx range, then using the same RNG state to shuffle the 162,165,686-entry shuffle_idx range. The complete regenerated array is exactly equal to this replacement.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment