datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
RecreationBench
RecreationBench
GitHub · Website · Leaderboard · Report ↗
RecreationBench is a benchmark of 250 application-recreation tasks across Ubuntu, macOS, Windows, Android, and Web. It is the held-out evaluation suite of RecreationWorld, a framework for scalable and verifiable environments for hybrid computer-use agents.
Given a high-level recreation request and interactive access to a running reference application, an agent explores its behavior, implements a new application, and… See the full description on the dataset page: https://huggingface.co/datasets/Qwen/RecreationBench.ccrawl-recrawl-domains
Common Crawl Domain Recrawl
Live fetches of the home page of every ranked domain in Common Crawl's web graph, rendered to Markdown as they are fetched
What is it?
Common Crawl's web graph ranks domains by how central they are, but it does not tell you what those domains actually serve today. This dataset walks that ranking from the top and fetches each domain's home page now, storing the response as one Parquet row with the body, the headers, the timing and the… See the full description on the dataset page: https://huggingface.co/datasets/open-index/ccrawl-recrawl-domains.Fanatic-Fandom
Dataset Card for Fanatic Fandom
Waifu to catch your attention.
Dataset Details
Dataset Description
Fanatic Fandom is a cleaned dataset of a raw scrape of fandom wikis. We crawled all the publicly available wikis and crawled each page.Filtering to a total amount of tokens of ~7.43B (llama-2-7b-chat-tokenizer) / ~6.27B (RWKV Tokenizer) from primarily English language.
Curated by: KaraKaraWitch
Funded by: Recursal.ai (I work there lol)
Shared by: KaraKaraWitch… See the full description on the dataset page: https://huggingface.co/datasets/recursal/Fanatic-Fandom.ReCITE
ReCITE: Real-world CausalIty from Textual Evidence Benchmark
ReCITE (Real-world CausalIty from Textual Evidence) is a benchmark for evaluating LLMs on causal graph extraction from real-world scientific text. It contains 292 annotated causal graphs from open-access MDPI and PLOS articles spanning diverse OpenAlex fields.
Paper: Can Large Language Models Infer Causal Relationships from Real-World Text?
GitHub: ReCITE Repository
Dataset Configurations
This dataset has… See the full description on the dataset page: https://huggingface.co/datasets/RyanSaklad/ReCITE.All-CVE-Records-Training-Dataset
CVE Chat‑Style Multi‑Turn Cybersecurity Dataset (1999 – 2025)
1. Project Overview
This repository hosts the largest publicly available chat‑style, multi‑turn cybersecurity dataset to date, containing ≈ 300 000 Common Vulnerabilities and Exposures (CVE) records published between 1999 and 2025. Each record has been meticulously parsed, enriched, and converted into a conversational format that is ideal for training and evaluating AI and AI‑Agent systems focused on… See the full description on the dataset page: https://huggingface.co/datasets/AlicanKiraz0/All-CVE-Records-Training-Dataset.ccrawl-recrawl-urls
Common Crawl URL Recrawl
Live refetches of pages from Common Crawl's URL index, with the body inline and the text already extracted
What is it?
Common Crawl publishes which URLs it saw and when, but the page bodies live in WARC archives that are awkward to query and are as old as the crawl that made them. This dataset takes the URL index for a single monthly crawl and fetches the pages again now, storing each response as one Parquet row with the body, the… See the full description on the dataset page: https://huggingface.co/datasets/open-index/ccrawl-recrawl-urls.Sequential-Math
RecursiveMAS Sequential-Math
Project Page | Code | Paper
We introduce RecursiveMAS, a multi-agent framework that scales agent collaboration through latent-space recursion. This dataset contains training examples for the Sequential-Style setting.
Dataset Details
Item
Description
Dataset
RecursiveMAS/Sequential-Math
Original file
Sequential-Math.json
Collaboration style
Sequential-Style
Used for
sequential math inner agents and outer RecursiveLink… See the full description on the dataset page: https://huggingface.co/datasets/RecursiveMAS/Sequential-Math.recube-data
Data
This directory contains all benchmark data for the Re2Code repository-level code reconstruction benchmark.
Download
All data files are hosted on Hugging Face and can be downloaded using:
# Install huggingface_hub if not already installed
pip install huggingface_hub
# Download the entire dataset
huggingface-cli download wlqmfl1999/recube-data --repo-type=dataset --local-dir data/
# Or download in Python
from huggingface_hub import snapshot_download… See the full description on the dataset page: https://huggingface.co/datasets/wlqmfl1999/recube-data.Europarl-Translation-Instruct
Dataset Card for Europarl-Translation-Instruct
Waifu to catch your attention.
Dataset Details
Dataset Description
europarl-translation-instruct is a translation instruct dataset built from europarl data.
Curated by: M8than
Funded by: Recursal.ai
Shared by: M8than
Language(s) (NLP): English instruct (but various languages in)
License: cc-by-sa-4.0
Dataset Sources
Source Data: https://www.statmt.org/europarl/ (Transcript source)
Processing… See the full description on the dataset page: https://huggingface.co/datasets/recursal/Europarl-Translation-Instruct.sequence-recovery
Next K-mer Prediction
Abouts
The Next K-mer Prediction task is a zero-shot evaluation method introduced in the GENERator paper to assess the quality of pretrained models. It involves inputting a sequence segment into the model and having it predict the next K base pairs. The predicted sequence is then compared to the actual sequence to assess accuracy.
Sequence: The input sequence has a maximum length of 96k base pairs (bp). You can control the number of input… See the full description on the dataset page: https://huggingface.co/datasets/GenerTeam/sequence-recovery.recursive-task-synthesis-glm-5.3-rollouts
GLM 5.3 agentic rollouts on Recursive-Task-Synthesis
This dataset catalogs the full collection made from the pinned
Recursive-Task-Synthesis dataset revision
be44f96808d5a9b599d5cb024341ff00091adeb7. The repository includes approximately 260.5 GiB of trajectory payload tar shards.
Contents at a glance
Item
Count
Source tasks considered
37,284
Source candidates inspected
19,368
Converted tasks after source filters
18,600
Tasks passing gold… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/recursive-task-synthesis-glm-5.3-rollouts.SuperWikiNEXT-32B
Dataset Card for SuperWikiNEXT-32B
Waifu to catch your attention.
Dataset Details
Dataset Description
SuperWikipedia-NEXT is an enhanced version of the SuperWIKI dataset. Which SuperWIKI was born out of the thought of a better filtered Wikipedia while retaining markdowns.
SuperWikipedia-NEXT contains ~32.44B Tokens (llama-2-7b-chat-tokenizer) / ~27.92B Tokens (RWKV Tokenizer) from approximately 60 "High quality" / "Selected" languages.
Curated by:… See the full description on the dataset page: https://huggingface.co/datasets/recursal/SuperWikiNEXT-32B.reprocessed_singapore_national_speech_corpus
Dataset Card for Reprocessed National Speech Corpus
NOTE: This is an Reprocessed version KaraKaraWitch from Recursal.The official download can be found here.
Dataset Details
Dataset Description
Dataset Description:
The National Speech Corpus (NSC) is the first large-scale Singapore English corpus, sponsored by the Info-communications and Media Development Authority (IMDA) of Singapore. The objective is to serve as a primary resource of open speech data for… See the full description on the dataset page: https://huggingface.co/datasets/recursal/reprocessed_singapore_national_speech_corpus.UltrachatBR
UltrachatBR: Um Dataset em Português baseado no Ultrachat
O UltrachatBR é uma versão em português do conhecido dataset Ultrachat, originalmente desenvolvido para o idioma inglês. Este projeto visa disponibilizar uma vasta coleção de diálogos traduzidos para o português, ampliando assim o acesso a recursos de processamento de linguagem natural para a comunidade de língua portuguesa.
Processo de Tradução
O processo de tradução foi realizado utilizando a API do Google… See the full description on the dataset page: https://huggingface.co/datasets/recogna-nlp/UltrachatBR.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 5.0000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 1.3000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 4.0000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think.recube-dataThis dataset serves as the official data repository for the ReCUBE benchmark; the corresponding evaluation codebase and execution instructions can be found at https://anonymous.4open.science/r/ReCUBE-E0FB/README.md.
Data
This directory contains all benchmark data for ReCUBE.
Download
All data files are hosted on Hugging Face and can be downloaded using:
# Install huggingface_hub if not already installed
pip install huggingface_hub
# Download the entire dataset… See the full description on the dataset page: https://huggingface.co/datasets/recube-anon-2026/recube-data.MDN
Dataset Card for MDN
Waifu to catch your attention.
Dataset Description
MDN is a ~57M Tokens (llama-2-7b-chat-tokenizer) / ~46.52M Tokens (RWKV Tokenizer) scrape of MDN (Developer.mozilla.org).
It serves as a training resource for large language models and other NLP tasks.
This card details the dataset's origin, content, and limitations.
Curated by:KaraKaraWitch
Funded by: Recursal.ai (I work there lol)
Shared by: KaraKaraWitch
Language(s) (NLP): English, Espanol… See the full description on the dataset page: https://huggingface.co/datasets/recursal/MDN.recon-eval
ReconEval — Financial Reconciliation Benchmark
Reading results from this benchmark. Four properties of ReconEval shape what
a score on it means. Anyone comparing models here should know them.
One class can dominate a margin. PARTIAL_MATCH is the highest-variance class
between models, and its 32 evaluation items are generated from 9 abbreviation
pairs — all of which also appear in the training split, overlap fraction 1.0. On
this set, "learned the concept" and "memorised nine… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/recon-eval.tool-reasoning-sft-CODING-text_to_terminal_v2-sft-tool-use-agent-data-cleaned-rectified
Text to Terminal, v2 — Cleaned & Rectified
👥 Follow the Author
Aman Priyanshu
Overview
This dataset is a cleaned, combined, and thinking-augmented version of muellerzr/text_to_terminal_v2. It pairs natural language instructions with their corresponding terminal/bash commands, now augmented with explicit <think> reasoning traces that model the step-by-step thought process before producing the final command.The restructuring approach is directly… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/tool-reasoning-sft-CODING-text_to_terminal_v2-sft-tool-use-agent-data-cleaned-rectified.EduBench
EduBench 📚
EduBench é um benchmark em português brasileiro para avaliação de Large Language Models (LLMs) em tarefas educacionais, composto por 3,149 questões discursivas extraídas de vestibulares de alta competitividade.
GitHub
Paper
Dataset Description
Fontes
USP: Universidade de São Paulo
UNICAMP: Universidade Estadual de Campinas
UNESP: Universidade Estadual Paulista
Período
2015-2025 (11 anos de provas)
Áreas do… See the full description on the dataset page: https://huggingface.co/datasets/recogna-nlp/EduBench.DriftSQL-Recovery
DriftSQL-Recovery
DriftSQL-Recovery is an execution-verified dataset for training and
evaluating SQL agents under schema, business-knowledge, and interaction drift.
It contains database-isolated recovery tasks, canonical seven-tool
trajectories, real on-policy failures, Recovery SFT examples, hard replay, and
full-episode GRPO records.
The accompanying implementation, data factory, sandbox, reward, and evaluation
code are available in DriftSQL.
Release scope… See the full description on the dataset page: https://huggingface.co/datasets/lxSYSU/DriftSQL-Recovery.CRAFT-RecipeGen
CRAFT-RecipeGen
This is a synthetic dataset generated with the CRAFT framework proposed in the paper CRAFT Your Dataset: Task-Specific Synthetic Data Generation Through Corpus Retrieval and Augmentation.
The correctness of the data has not been verified in detail.
4 synthetic dataset sizes (S, M, L, XL) are available.
Compared to other synthetically generated datasets with the CRAFT framework, this task did not scale similarly well and we do not match the performance of general… See the full description on the dataset page: https://huggingface.co/datasets/ingoziegler/CRAFT-RecipeGen.VeriLoop-Governed-Recurrence-Verified
VLR-Recurrence-Verified
VLR-Recurrence-Verified is a synthetic-data construction release for studying
evidence-convergent program repair. It operationalizes a protected partial order:
a candidate is positive only when it preserves every already-satisfied
obligation and strictly improves at least one unresolved obligation.
Scale
Split
Tasks
Families
Transitions
Balanced pairs
Certified finals
Train
3,500
28
12,250
49,000
3,500
Validation
750
10
2,623… See the full description on the dataset page: https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-Governed-Recurrence-Verified.RecurrReason
RecurrReason: Recurrent Reasoning on Symbolic Puzzles
A difficulty-controlled benchmark for evaluating multi-step reasoning in language models
📋 Table of Contents
Overview
Dataset Structure
Puzzles
Quick Start
Citation
License
🎯 Overview
RecurrReason is a benchmark of four recurrent logic puzzles with optimal trajectories and controlled difficulty scaling (N=1 to 10). It tests whether language models can:
Find optimal (minimal-length)… See the full description on the dataset page: https://huggingface.co/datasets/gmannem/RecurrReason.food-recipes
Food.com Multimodal Recipe Dataset (15K)
Dataset Summary
Property
Value
Total samples
~15,000
Modalities per sample
2 — PNG recipe card image + Markdown text
Image format
PNG, 300 DPI, A4 aspect ratio
Source dataset
Food.com Recipes and User Interactions (Kaggle)
Raw recipe pool
~231,637 recipes
License
See source dataset license
Intended Use Cases
This dataset was designed to support the following downstream research and engineering… See the full description on the dataset page: https://huggingface.co/datasets/rahul7star/food-recipes.pubmedqa-recursive-llm-degradation-qwen2.5-0.5b
PubMedQA Recursive LLM Degradation — Qwen2.5-3B
This repository contains synthetic biomedical question-answering data
and model predictions generated as part of a study of recursive
fine-tuning and model degradation.
Base Model
Qwen/Qwen2.5-3B
Source Dataset
The experiments use the PubMedQA dataset:
qiaoxin/PubMedQA
This repository contains generated/derived research artifacts and does
not redistribute the original PubMedQA dataset in its entirety.… See the full description on the dataset page: https://huggingface.co/datasets/chrislimbe/pubmedqa-recursive-llm-degradation-qwen2.5-0.5b.gigaverbo-v2-rec-sft
GigaVerbo-v2 REC SFT
A model should not merely know how to reason; it should learn when reasoning is worth the cost.
Dataset repository: OliveiraJLT/gigaverbo-v2-rec-sftBase dataset: Polygl0t/gigaverbo-v2-sftAnswer-generation model: openai/gpt-oss-20bQuality classifier: Polygl0t/portuguese-qwen3-4b-instruct-quality-classifierReasoning translation model and token accounting tokenizer: Qwen/Qwen3.5-9B
Dataset Summary
GigaVerbo-v2 REC SFT — short for GigaVerbo-v2… See the full description on the dataset page: https://huggingface.co/datasets/OliveiraJLT/gigaverbo-v2-rec-sft.pubmedqa-recursive-llm-degradation-qwen2.5-3b
PubMedQA Recursive LLM Degradation — Qwen2.5-3B
This repository contains synthetic biomedical question-answering data
and model predictions generated as part of a study of recursive
fine-tuning and model degradation.
Base Model
Qwen/Qwen2.5-3B
Source Dataset
The experiments use the PubMedQA dataset:
qiaoxin/PubMedQA
This repository contains generated/derived research artifacts and does
not redistribute the original PubMedQA dataset in its entirety.… See the full description on the dataset page: https://huggingface.co/datasets/chrislimbe/pubmedqa-recursive-llm-degradation-qwen2.5-3b.