Delta-Ngram experiment data

Frozen data inputs and reproduction metadata for the Delta-Ngram experiments. The companion writer weights and loading code are in haowu11/Delta-Ngram (model repository).

Contents

Directory	Contents
five_source_train/	The code, papers, books, web and news corpus, with explicit train/validation membership, source metadata and compressed native token caches.
dev64/	The fixed DEV64 evaluation articles, provenance, event-pair registry and native evaluation inputs.
seven_tasks/	The seven morphology tasks: 35,000 training examples, 448 fixed evaluation examples, task order, scoring protocol and source reconstruction tools.

The ordinary-document corpus contains 38,566 training documents and 3,069 validation documents. Its additional news-event sampling stream contains 713 training streams and 224 validation streams, backed by 2,002 article records. The frozen native index therefore contains 43,637 article records in total: 40,083 train and 3,554 validation. News events use the same news source; stream counts and article counts describe different units. The component metadata gives exact membership and counts.

DEV64 contains 64 event families, backed by 146 source articles. Its five unrelated-article distances are 0, 1, 4, 8 and 16. Family is the paired resampling unit; repeated distances and executions of a family do not create additional independent families.

The seven-task order is sme → nav → hye → afb → sqi → mkd → amh, with 5,000 training examples and 64 evaluation examples per task. Preserve the stored example order, identifiers, prompts, targets and alternative references when reproducing the experiment.

Download and verify

Authenticate with a Hugging Face account that has access to this private repository, then download the dataset repository:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="haowu11/Delta-Ngram",
    repo_type="dataset",
    local_dir="Delta-Ngram-data",
)

Verify all published files without loading a model or using a GPU:

python Delta-Ngram-data/verify_dataset.py

release_manifest.json records the path, size and SHA256 of every release member except itself. The upload receipt separately records the manifest hash and checks the committed files by downloading them again. Component manifests preserve the source identities and preparation details.

Read each component's README before restoring caches or running evaluation. Raw text, sampling streams, token caches and evaluation plans have different schemas. Load the desired component explicitly rather than treating all files in the repository as a single table. Historical manifests may preserve original machine paths for provenance; use the documented restoration paths for a new machine.

Reproduction scope

The recorded data versions, sample membership and preprocessing choices are part of the experiment. Use the specified tokenizer/model revisions and special-token conventions. Token caches are tied to their recorded tokenizer identity. Training's approximately 100 million cross-entropy targets measure consumed training targets, not the number of unique words or unique raw tokens in this corpus.

This release provides data and protocols. Evaluation results and ongoing experiment outputs are maintained separately. Keep validation, DEV64 and seven-task evaluation records out of the writer's training data.

Sources and licenses

The source records retain their upstream licenses, attribution and provenance. The repository-level mixed-source-licenses label describes this collection and does not relicense upstream text. In particular, the seven UniMorph language datasets retain CC BY-SA 3.0 and their accompanying notices. See LICENSES.md and the component provenance and license files for the applicable terms and pinned sources.

Original packaging and verification code authored for this release is provided under Apache-2.0; included third-party material retains its own terms.