waylonli/Selective-Attention-Freezing-data

Selective Attention Freezing: Data and Splits

Evaluation data and partition records for When Can Attention Heads Be Statically Defined? by Weixian Waylon Li, Yintao Tai, Marcio Fonseca and Shay B. Cohen. Code and evaluation commands.

Directory	Contents
pretraining/fineweb_edu/	Exact archived validation tokens for the native 124M checkpoints, plus split metadata.
pretraining/fineweb_edu_20b/	Exact archived validation tokens for the native 1B checkpoints, plus the pinned training-data revision and hashes.
downstream/	Zero-based train, internal selection-dev and official validation indices for SST-2, BoolQ and QuALITY, seeds 1337--1339.
mqar/	Six exact MQAR evaluation sets at 512 tokens and 8, 16, 24, 32, 48 or 64 key-value pairs.

SHA256SUMS.json records file checksums. Both val.bin files contain little-endian uint16 GPT-2 token IDs. Their first halves are the controller partitions; their second halves are used for the reported native-checkpoint PPL evaluation. Use the token ranges in each manifest, then form non-overlapping windows at the checkpoint's native context length. The maturity study uses its separate data protocol and should not be evaluated as though it used these native-checkpoint partitions.

The downstream index files contain no articles, questions or answers. Load the original datasets identified in each file, verify their recorded fingerprints, and select the specified row indices from the indicated source split. These partitions were reconstructed from the cached datasets and checked against the completed experiment records.

MQAR arrays can be read with numpy.load(path, allow_pickle=False); inputs and labels are int64 arrays of shape (512, 512). Labels equal to -100 are ignored. Input and label hashes match the archived pair-sweep experiments. The manifest also specifies the deterministic training-data generator.

Training token files and model weights are not part of this dataset repository. The 1B training file can be regenerated with the revision-pinned preparation command in the code README. The original upstream revision for the 124M training corpus is still being recovered; the provided 124M validation file is the actual archived file, not a newly sampled substitute.

Sources and Licences

FineWeb-Edu validation tokens are derived from HuggingFaceFW/fineweb-edu, released under ODC-By 1.0. Attribute the FineWeb-Edu creators and preserve the upstream notices and applicable terms when redistributing the token files. The source material may have additional rights and terms as described in the upstream dataset card.

Downstream data remain with GLUE SST-2, SuperGLUE BoolQ and QuALITY; this repository supplies partition indices only. MQAR data are generated using the Zoology generator, whose Apache-2.0 licence and attribution are retained in the code repository.