datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
swallow-math-v2
SwallowMath-v2
Resources
📑 arXiv: Read our paper for detailed methodology at arXiv:2505.02881.
🤗 Sister Dataset: Discover SwallowCode2, our companion dataset for code generation.
🧮 What is it?
SwallowMath-v2 is a large-scale mathematical dataset containing 32 billion tokens, developed as the successor to SwallowMath-v1.
Building on the success of v1, this release aims to construct a larger-scale and more permissively licensed corpus to support open and… See the full description on the dataset page: https://huggingface.co/datasets/tokyotech-llm/swallow-math-v2.swallow-code-v2
SwallowCode-v2
Resources
📑 arXiv: Read our paper for detailed methodology and results at arXiv:2505.02881.
🤗 Sister Dataset: Discover SwallowMath-v2, our companion dataset for mathematical reasoning.
💻 What is it?
SwallowCode-v1 was a high-quality Python code dataset generated through an LLM-based rewriting pipeline.
However, it had two significant limitations:
(1) it was distributed under the Llama 3.3 Community License, and
(2) its size was limited to… See the full description on the dataset page: https://huggingface.co/datasets/tokyotech-llm/swallow-code-v2.tokenized_datasetcode_instructions_122k_alpaca_styleEmilia-Dataset-tokenisedtokenized_C4temp-decoder-train-tokenizedmeta-llama-Llama-3.2-1B-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
M-IFEval-Jaswallow-math
SwallowMath
October 21, 2025: Newer versions are available: SwallowCode-v2 and SwallowMath-v2 have been released with improved rewriting pipelines.
Resources
🐙 GitHub: Explore the project repository, including pipeline code and prompts at rioyokotalab/swallow-code-math.
📑 arXiv: Read our paper for detailed methodology and results at arXiv:2505.02881.
🤗 Sister Dataset: Discover SwallowCode, our companion dataset for code generation.
What is it?… See the full description on the dataset page: https://huggingface.co/datasets/tokyotech-llm/swallow-math.swallow-code
SwallowCode
Notice
May 21, 2025: We have deleted ablation/exp1-the-stack-v2-train-smol-ids-python because it was flagged as potentially containing unsafe data collected from the Python subset of https://huggingface.co/datasets/bigcode/the-stack-v2-train-smol-ids. However, since this dataset can be reconstructed from the-stack-v2-train-smol-ids, there is no issue in terms of reproducibility.
May 21, 2025: ClamAV has flagged “Win.Trojan.MSShellcode-88” in… See the full description on the dataset page: https://huggingface.co/datasets/tokyotech-llm/swallow-code.tokenized-dataset-combineSwallow-Nemotron-Post-Training-Dataset-v1
Swallow-Nemotron-Post-Training-Dataset-v1
The Swallow LLM Project constructed the Swallow-Nemotron-Post-Training-Dataset-v1 based on the math, code, and stem subsets of the NVIDIA Nemotron-Post-Training-Dataset-v1, as illustrated in the figure below.
Dataset Construction
The original Thinking Trajectories and Assistant Outputs in the Nemotron-Post-Training-Dataset-v1 were synthesized using DeepSeek-R1-0528.
However, we identified an issue with the Thinking… See the full description on the dataset page: https://huggingface.co/datasets/tokyotech-llm/Swallow-Nemotron-Post-Training-Dataset-v1.aya-expanse-8b-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
gpt-4o-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
DynamicMCPBench
DynamicMCPBench
A trace-grounded, effect-scored benchmark for LLM agents on live MCP servers.
Tasks are generated forward: an explorer agent drives real MCP tools until a goal
is reached, the recorded trace is distilled into a TaskSpec, and candidates are graded
on whether they reproduce the effects the trace produced — checkpoints, equivalence
sets, minefields, a partial order — never on matching an answer string or a fixed tool
list. Candidates are evaluated under… See the full description on the dataset page: https://huggingface.co/datasets/TokenWasteGroup/DynamicMCPBench.tokenised_subsetof_erickfmm__red_pajama_es_hq_35urls-tokenized
URLs (tokenized)
ks46/urls-sampled run through a byte-level
BPE built for URLs, stored as flat uint16 token streams that memory-map
directly into a training loop.
Shards
512
URLs
18,729,786,698
Tokens
664,731,047,208
Vocabulary
8,192
Token dtype
uint16, little-endian
There is no parquet here and the dataset viewer will not render it. These
are raw token bins; see Reading the data below.
Layout
tokenizer/ the exact vocabulary… See the full description on the dataset page: https://huggingface.co/datasets/ks46/urls-tokenized.jora_corpus1_FR_tokenized_128ktokenizedfacebook-xglm-564M-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
sokoban-10k-vjepa2-tokenizedQwen3.8-27B-Distill-1M-3.12B-Tokens
Qwen3.8-27B-Distill-1M-4.83B-Tokens
A unified, globally deduplicated, large-scale supervised distillation corpus built from 992,318 conversations generated by Qwen/Qwen3.8-27B, containing 4,834,771,862 target output tokens (3,570,459,498 reasoning tokens + 1,264,312,364 final response tokens) and 5,104,980,053 total sequence tokens.
1. Dataset Overview
This dataset merges, aligns, and deduplicates the two primary high-quality Qwen3.8-27B generation corpora on… See the full description on the dataset page: https://huggingface.co/datasets/MaxDevv/Qwen3.8-27B-Distill-1M-3.12B-Tokens.fineweb_bltQwen-Qwen3-8B-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
brand-heavy-token-quality-datasetbyt5-small-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture
Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value while the traces are ~3x longer, by keeping only a subset of the expanded rows. Its sibling arm holds the ROW share instead; together they separate more deliberation from more difficult-advice signal.
field
value
experiment
Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture.general-product-token-quality-datasetstructure-heavy-token-quality-dataset