quinnlue/regmix-data-sample-stella

RegMix Data Sample (Stella Embeddings)

Mirror of sail/regmix-data-sample with precomputed 1024-dim L2-normalized embeddings from Marqo/dunzhang-stella_en_400M_v5.

Dataset Description

This dataset preserves the original RegMix domain JSONL text and adds Stella embeddings organized for downstream clustering and mixture analysis.

• Source: sail/regmix-data-sample
• Embedding model: Marqo/dunzhang-stella_en_400M_v5
• Embedding dim: 1024
• Max token length: 512
• Pooling: mean over token hidden states
• Normalization: L2
• Total rows: 2,345,364

Domains

• train/arxiv: 79,387 rows
• train/dm_mathematics: 64,147 rows
• train/enron_emails: 30,905 rows
• train/europarl: 4,348 rows
• train/freelaw: 169,944 rows
• train/github: 300,000 rows
• train/gutenberg_pg_19: 2,276 rows
• train/hackernews: 52,544 rows
• train/nih_exporter: 58,972 rows
• train/philpapers: 2,024 rows
• train/pile_cc: 300,000 rows
• train/pubmed_abstracts: 300,000 rows
• train/pubmed_central: 150,000 rows
• train/stackexchange: 84,651 rows
• train/ubuntu_irc: 623 rows
• train/uspto_backgrounds: 300,000 rows
• train/wikipedia_en: 265,170 rows
• validation/arxiv: 2,406 rows
• validation/dm_mathematics: 1,922 rows
• validation/enron_emails: 1,010 rows
• validation/europarl: 157 rows
• validation/freelaw: 5,101 rows
• validation/github: 18,195 rows
• validation/gutenberg_pg_19: 80 rows
• validation/hackernews: 1,632 rows
• validation/nih_exporter: 1,884 rows
• validation/philpapers: 64 rows
• validation/pile_cc: 52,790 rows
• validation/pubmed_abstracts: 29,895 rows
• validation/pubmed_central: 5,911 rows
• validation/stackexchange: 30,378 rows
• validation/ubuntu_irc: 22 rows
• validation/uspto_backgrounds: 11,415 rows
• validation/wikipedia_en: 17,511 rows

Schema

Each parquet shard under data/{split}/{domain}.parquet contains:

Column	Type	Description
example_id	int64	Stable join key
split	string	train or validation
domain	string	RegMix domain name
source_file	string	Original JSONL filename
line_idx	int32	Line index in source file
text	string	Original document text
text_char_len	int32	Character length of text
embedding	list[float32]	L2-normalized Stella vector

indices/global_index.parquet maps each example_id to its shard path and row index.

Usage

Load with Hugging Face Datasets

from datasets import load_dataset
import numpy as np

ds = load_dataset("quinnlue/regmix-data-sample-stella", split="train")
emb = np.stack(ds["embedding"], dtype=np.float32)  # [N, 1024]
ids = np.array(ds["example_id"], dtype=np.int64)
domains = ds["domain"]

Load a single domain

from datasets import load_dataset

arxiv = load_dataset(
    "quinnlue/regmix-data-sample-stella",
    data_files={"train": "data/train/arxiv.parquet"},
    split="train",
)

Clustering starter (spherical k-means)

from datasets import load_dataset
import numpy as np
from sklearn.cluster import KMeans

ds = load_dataset("quinnlue/regmix-data-sample-stella", split="train")
emb = np.stack(ds["embedding"], dtype=np.float32)
ids = np.array(ds["example_id"], dtype=np.int64)

kmeans = KMeans(n_clusters=128, n_init=10, random_state=0)
cluster_ids = kmeans.fit_predict(emb)

assignments = {int(example_id): int(cluster_id) for example_id, cluster_id in zip(ids, cluster_ids)}

Citation

If you use this dataset, please cite the RegMix paper:

@article{liu2024regmix,
  title={RegMix: Data Mixture as Regression for Language Model Pre-training},
  author={Liu, Qian and Zheng, Xiaosen and Muennighoff, Niklas and Zeng, Guangtao and Dou, Longxu and Pang, Tianyu and Jiang, Jing and Lin, Min},
  journal={arXiv preprint arXiv:2407.01492},
  year={2024}
}