lucky-verma/dyt-composition-artifacts

DyT Composition Study Artifacts

This dataset contains sanitized result manifests and analysis outputs for When Does Removing LayerNorm Help? Activation Bounding as a Regime-Dependent Implicit Regularizer.

DOI: https://doi.org/10.48550/arXiv.2604.23434

Contents

The artifacts include aggregate training metrics, saturation measurements, statistical-test summaries, predictor-validation outputs, table-source manifests, and selected aggregate analysis files used by the public code repository.

The Dataset Viewer table is an index of the artifact files. The machine-readable result artifacts are stored under results/.

This is not a natural-language training dataset. It does not redistribute WikiText, OpenWebText, LAMBADA, BLIMP, model checkpoints, or raw training logs.

Intended Use

Use this artifact bundle to:

• inspect the machine-readable results behind the paper;
• reproduce paper tables and consistency checks;
• compare DyT, LayerNorm, RMSNorm, HardTanh, DiffAttn, and related controls at the reported scales;
• audit provenance for reported quantitative claims.

Limitations

• The experiments are compute-limited and below Chinchilla-optimal training.
• The included files are result artifacts, not full raw training traces or checkpoints.
• The saturation diagnostic should be treated as a per-deployment calibration cue, not a universal rule.
• Raw public datasets retain their original licenses and are not mirrored here.

Licensing

The result artifacts in this dataset are released under CC BY 4.0.

The associated GitHub code is released under the MIT License.

Citation

@misc{verma2026dytcomposition,
  title        = {When Does Removing LayerNorm Help? Activation Bounding as a Regime-Dependent Implicit Regularizer},
  author       = {Verma, Lucky},
  year         = {2026},
  publisher    = {arXiv},
  doi          = {10.48550/arXiv.2604.23434},
  url          = {https://arxiv.org/abs/2604.23434},
  eprint       = {2604.23434},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG}
}