Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
text
string
prompt
string
completion
string
"Chapter: _5 May._--I must have been asleep, for certainly if I had been fully\nawake I must have no(...TRUNCATED)
"Chapter: _5 May._--I must have been asleep, for certainly if I had been fully\nawake I must have no(...TRUNCATED)
" Dracula's castle is described, like almost everything else, in precise detail. Harker notes the ca(...TRUNCATED)
"Chapter: _12 September._--How good they all are to me. I quite love that dear Dr.\nVan Helsing. I w(...TRUNCATED)
"Chapter: _12 September._--How good they all are to me. I quite love that dear Dr.\nVan Helsing. I w(...TRUNCATED)
" On the 12th of September, Lucy is perplexed by the presence of the garlic flowers, but she has suc(...TRUNCATED)
"Chapter: _23 September_.--Jonathan is better after a bad night. I am so glad that\nhe has plenty of(...TRUNCATED)
"Chapter: _23 September_.--Jonathan is better after a bad night. I am so glad that\nhe has plenty of(...TRUNCATED)
" Mina decides to transcribe the journal which Jonathan kept at the Castle Dracula in Transylvania. (...TRUNCATED)
"Chapter: When we arrived at the Berkeley Hotel, Van Helsing found a telegram\nwaiting for him:--\n\(...TRUNCATED)
"Chapter: When we arrived at the Berkeley Hotel, Van Helsing found a telegram\nwaiting for him:--\n\(...TRUNCATED)
" Dr. Seward's diary continues sometime later, and he details for us his first meeting with Mina Har(...TRUNCATED)
"Chapter: _1 October, evening._--I found Thomas Snelling in his house at Bethnal\nGreen, but unhappi(...TRUNCATED)
"Chapter: _1 October, evening._--I found Thomas Snelling in his house at Bethnal\nGreen, but unhappi(...TRUNCATED)
" Jonathan, through his persistent investigations, discovers the whereabouts of twelve more of the b(...TRUNCATED)
"Chapter: This to Jonathan Harker.\n\nYou are to stay with your dear Madam Mina. We shall go to make(...TRUNCATED)
"Chapter: This to Jonathan Harker.\n\nYou are to stay with your dear Madam Mina. We shall go to make(...TRUNCATED)
" Van Helsing thinks that Jonathan Harker should stay in England with his wife, since he now knows t(...TRUNCATED)
"Chapter: _29 October._--This is written in the train from Varna to Galatz. Last\nnight we all assem(...TRUNCATED)
"Chapter: _29 October._--This is written in the train from Varna to Galatz. Last\nnight we all assem(...TRUNCATED)
" On the 29th day of October, Dr. Seward records that Mina, under hypnosis, can hear and distinguish(...TRUNCATED)
"Chapter: As I rose and dressed, I thought over what had happened, and wondered if\nit were a dream.(...TRUNCATED)
"Chapter: As I rose and dressed, I thought over what had happened, and wondered if\nit were a dream.(...TRUNCATED)
" The next morning, Jane wakes, wondering if the previous night was just a dream. She feels transfor(...TRUNCATED)
"Chapter: A week passed, and no news arrived of Mr. Rochester: ten days, and still\nhe did not come.(...TRUNCATED)
"Chapter: A week passed, and no news arrived of Mr. Rochester: ten days, and still\nhe did not come.(...TRUNCATED)
" Jane cannot help missing Rochester and she longs for his return. A letter is received giving the s(...TRUNCATED)
"Chapter: Some time in the afternoon I raised my head, and looking round and seeing\nthe western sun(...TRUNCATED)
"Chapter: Some time in the afternoon I raised my head, and looking round and seeing\nthe western sun(...TRUNCATED)
" It is nearly the end of the day before Jane emerges from her room, but she is already resolved to (...TRUNCATED)
End of preview. Expand in Data Studio

DHSA_Long-Data-Collections

A length-bucketed release of togethercomputer/Long-Data-Collections, used in Long-Context Modeling with Dynamic Hierarchical Sparse Attention for Memory-Constrained LLM Inference (ICML 2026 Spotlight).

Each example is assigned to exactly one length bucket based on token count with meta-llama/Llama-3.1-8B-Instruct (add_special_tokens=True):

Bucket Token range
lt_8k [0, 8K)
8k_16k [8K, 16K)
16k_32k [16K, 32K)
32k_64k [32K, 64K)
64k_128k [64K, 128K)
gt_128k β‰₯ 128K
  • Pretrain examples are bucketed by the text field.
  • Fine-tune examples are bucketed by the prompt field.

Directory layout

Long-Data-Collections/
β”œβ”€β”€ README.md
β”œβ”€β”€ length_bucket_summary.json
β”œβ”€β”€ pretrain/
β”‚   β”œβ”€β”€ lt_8k/
β”‚   β”‚   β”œβ”€β”€ arxiv_doc_to_abs.jsonl.zst
β”‚   β”‚   β”œβ”€β”€ NI_decontaminated_materialized.jsonl.zst
β”‚   β”‚   β”œβ”€β”€ P3_decontaminated_materialized.jsonl.zst
β”‚   β”‚   β”œβ”€β”€ pile_sub.jsonl.zst
β”‚   β”‚   β”œβ”€β”€ rp_sub.jsonl.zst
β”‚   β”‚   └── ul2_plus_oscar_en.jsonl.zst
β”‚   β”œβ”€β”€ 8k_16k/
β”‚   β”œβ”€β”€ 16k_32k/
β”‚   β”œβ”€β”€ 32k_64k/
β”‚   β”œβ”€β”€ 64k_128k/
β”‚   └── gt_128k/
└── fine-tune/
    β”œβ”€β”€ lt_8k/
    β”‚   β”œβ”€β”€ booksum.jsonl.zst
    β”‚   └── natural_questions_10_200_docs.jsonl.zst
    β”œβ”€β”€ 8k_16k/
    β”œβ”€β”€ 16k_32k/
    β”œβ”€β”€ 32k_64k/
    β”œβ”€β”€ 64k_128k/
    └── gt_128k/

All files are zstd-compressed JSONL (.jsonl.zst). Each line is a JSON object preserving the original fields (text, and optionally meta / metadata for pretrain; text, prompt, completion for fine-tune).

Example counts by bucket

Pretrain (bucketed by text length)

Dataset lt_8k 8k_16k 16k_32k 32k_64k 64k_128k gt_128k Total
arxiv_doc_to_abs 377,854 559,745 436,259 150,632 28,557 5,258 1,558,305
rp_sub 913,818 11,265 3,893 1,019 368 151 930,514
ul2_plus_oscar_en 3,148,215 342,739 60,306 9,213 1,804 236 3,562,513
pile_sub 1,878,091 43,443 12,818 3,022 1,799 1,285 1,940,458
NI_decontaminated_materialized 460,066 117,188 567 210 12 0 578,043
P3_decontaminated_materialized 757,542 56,465 8 0 0 0 814,015
All pretrain 7,535,586 1,130,845 513,851 164,096 32,540 6,930 9,383,848

Fine-tune (bucketed by prompt length)

Dataset lt_8k 8k_16k 16k_32k 32k_64k 64k_128k gt_128k Total
booksum 7,909 1,207 413 64 5 2 9,600
natural_questions_10_200_docs 21,370 25,879 41,639 69 0 0 88,957
All fine-tune 29,279 27,086 42,052 133 5 2 98,557

Full counts are also available in length_bucket_summary.json.

Dataset description

This collection compiles long-context datasets for training and evaluating models that require extensive comprehension over large text inputs. It is derived from the public Together AI release and reorganized into length buckets for long-context experiments.

Pretrain data (pretrain/)

File Description
rp_sub RedPajama book subset β€” diverse literary text
arxiv_doc_to_abs RedPajama ArXiv papers with abstract appended after the paper body
ul2_plus_oscar_en UL2-style fill-in-the-blank completions (LAION Open-Instruction-Generalist)
pile_sub Subsample of The Pile
NI_decontaminated_materialized Natural Instructions, decontaminated against HELM core scenarios
P3_decontaminated_materialized Public Pool of Prompts (P3), decontaminated against HELM core scenarios

Fine-tune data (fine-tune/)

File Description
natural_questions_10_200_docs Multi-passage QA from Natural Questions (10–200 Wiki passages per question)
booksum Long-context book summarization

Fine-tune records contain prompt (context + instruction), completion (target answer/summary), and text (prompt + completion).

Usage

Load a specific bucket and file with the Hugging Face datasets library:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="pretrain/16k_32k/arxiv_doc_to_abs.jsonl.zst",
    split="train",
)

Or stream locally with zstd:

import json, zstandard as zstd

path = "pretrain/8k_16k/rp_sub.jsonl.zst"
with open(path, "rb") as f:
    with zstd.ZstdDecompressor().stream_reader(f) as reader:
        for line in reader:
            row = json.loads(line)
            ...

Licensing

Please refer to the original sources of each sub-dataset for their respective licenses. This dataset is a length-bucketed release of togethercomputer/Long-Data-Collections; upstream licensing terms apply.

Citation

@article{xiong2025long,
  title={Long-Context Modeling with Dynamic Hierarchical Sparse Attention for On-Device LLMs},
  author={Xiong, Siheng and Zou, Joe and Fekri, Faramarz and Cho, Yae Jee},
  journal={arXiv preprint arXiv:2510.24606},
  year={2025}
}
Downloads last month
96

Collection including sxiong/DHSA_Long-Data-Collections

Paper for sxiong/DHSA_Long-Data-Collections