bergson-wikitext-2-4656-chunks

Pre-tokenized WikiText-2 in fixed 512-token GPT-2 chunks, used by the bergson WikiText-2 / GPT-2 data-attribution replication of Bae et al. 2024 (Training Data Attribution via Approximate Unrolled Differentiation).

Each row is one 512-token document (one block of the concatenated corpus).

Split	Rows
train	4656
validation	481

Columns:

• input_ids — list[int32], length 512 (GPT-2 BPE token ids)
• length — int64, always 512

Load with chunk_length: 0 in a bergson config (the data is already tokenized and chunked).

How this was produced

The standard HuggingFace run_clm preprocessing recipe — the same one kronfluence's examples/wikitext uses — applied to public Salesforce/wikitext wikitext-2-raw-v1:

1. Tokenize the raw text column with the GPT-2 tokenizer, no added special tokens.
2. group_texts: within each 1000-row datasets.map batch, concatenate the tokens, drop the remainder, and slice into fixed 512-token blocks.

The 1000-row batching is load-bearing: it is what yields exactly 4656 / 481 chunks (whole-corpus concatenation instead gives 4671 / 482). No shuffling or seed is involved — the recipe is deterministic.

from datasets import DatasetDict, load_dataset
from transformers import AutoTokenizer

BLOCK = 512
tok = AutoTokenizer.from_pretrained("gpt2")
raw = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1")


def group_texts(examples):
    concat = sum(examples["input_ids"], [])
    total = (len(concat) // BLOCK) * BLOCK
    return {"input_ids": [concat[i : i + BLOCK] for i in range(0, total, BLOCK)]}


out = {}
for split in ("train", "validation"):
    toks = raw[split].map(
        lambda b: tok(b["text"]),
        batched=True,
        remove_columns=raw[split].column_names,
    )
    blocks = toks.map(
        group_texts,
        batched=True,
        remove_columns=[c for c in toks.column_names if c != "input_ids"],
    )
    out[split] = blocks.map(lambda x: {"length": len(x["input_ids"])})

DatasetDict(out).push_to_hub("EleutherAI/bergson-wikitext-2-4656-chunks")

This is also committed as examples/replication/prep_dataset.py in the bergson repo. The chunking reproduces the reference dataset byte-for-byte (identical input_ids on all 4656 train and 481 validation rows).

License

Inherits WikiText-2's CC BY-SA 3.0.