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	≥ 128K

• Pretrain examples are bucketed by the text field.
• Fine-tune examples are bucketed by the prompt field.

Directory layout

Long-Data-Collections/
├── README.md
├── length_bucket_summary.json
├── pretrain/
│   ├── lt_8k/
│   │   ├── arxiv_doc_to_abs.jsonl.zst
│   │   ├── NI_decontaminated_materialized.jsonl.zst
│   │   ├── P3_decontaminated_materialized.jsonl.zst
│   │   ├── pile_sub.jsonl.zst
│   │   ├── rp_sub.jsonl.zst
│   │   └── ul2_plus_oscar_en.jsonl.zst
│   ├── 8k_16k/
│   ├── 16k_32k/
│   ├── 32k_64k/
│   ├── 64k_128k/
│   └── gt_128k/
└── fine-tune/
    ├── lt_8k/
    │   ├── booksum.jsonl.zst
    │   └── natural_questions_10_200_docs.jsonl.zst
    ├── 8k_16k/
    ├── 16k_32k/
    ├── 32k_64k/
    ├── 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 — 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}
}