epfl-dlab/zip2zip-wikitext-repeat-phi35

Zip2Zip Repeated WikiText Stress Tests (Phi-3.5)

This repository contains two controlled evaluation corpora for studying merge-size transfer in Zip2Zip models. They are derived from the document-level WikiText-2 raw test split and built specifically with the microsoft/Phi-3.5-mini-instruct tokenizer.

Configurations

Config	Repetitions per source block	Rows	Repeated base tokens	SHA-256 of test.jsonl
repeat4	4	1,329	1,297,250	82fce0f7667d48f83f627ebc14b8148c64311127f38bc66741986135512a4c74
repeat8	8	2,651	2,617,485	48aa4185c6c961763ebea17ea8e46bb41eb8fd63300384643dd0ab2af2d3b049

Load either configuration with Hugging Face Datasets:

from datasets import load_dataset

repeat4 = load_dataset(
    "epfl-dlab/zip2zip-wikitext-repeat-phi35",
    "repeat4",
    split="test",
)
repeat8 = load_dataset(
    "epfl-dlab/zip2zip-wikitext-repeat-phi35",
    "repeat8",
    split="test",
)

Construction

We start from EleutherAI/wikitext_document_level, configuration wikitext-2-raw-v1, split test, and apply the standard WikiText-2 detokenization used by lm-evaluation-harness. Each document is tokenized with Phi-3.5, divided into token-aware source blocks, decoded, and repeated either four or eight times with blank-line separators. Blocks are shortened as needed so that every resulting row contains at most 1,000 Phi-3.5 base tokens and fits within a single 1,024-token evaluation window. Every repeated copy is included in the language-modeling loss.

The JSONL rows contain the repeated text, UTF-8 byte and whitespace-delimited word denominators, source document and block identifiers, and token-count metadata. The accompanying manifests record aggregate construction statistics.

Intended use

Repeat-4 and Repeat-8 are opt-in stress tests for hierarchical hyper-token merge-size transfer. Repetition increases the frequency of long LZW dictionary entries, making the difference between maximum merge sizes 3 and 4 measurable. They are not general-purpose language-modeling benchmarks. Later repeated copies are intentionally easier to predict, so absolute perplexities must not be compared across standard WikiText, Repeat-4, and Repeat-8. Model comparisons should be made within the same configuration and evaluation segmentation.

Because row boundaries are chosen using Phi-3.5 tokenization, these artifacts should not be silently reused with another tokenizer. Supporting another tokenizer requires rebuilding and publishing separately identified artifacts.

Reproducibility

The source code used to construct the corpora is scripts/build_repeated_wikitext.py in the Zip2Zip repository. The checked-in Zip2Zip lm-eval tasks are named zip2zip_wikitext_repeat4 and zip2zip_wikitext_repeat8.

Source and license

The source corpus is the WikiText-2 raw test split from EleutherAI/wikitext_document_level, which is distributed under the Creative Commons Attribution-ShareAlike 3.0 license. These transformed evaluation artifacts are released under the same license. Users should also cite the original WikiText paper:

@inproceedings{merity2017pointer,
  title     = {Pointer Sentinel Mixture Models},
  author    = {Stephen Merity and Caiming Xiong and James Bradbury and Richard Socher},
  booktitle = {International Conference on Learning Representations},
  year      = {2017}
}