datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
JSONSchemaBench
JSONSchemaBench
JSONSchemaBench is a benchmark of real-world JSON schemas designed to evaluate structured output generation for Large Language Models (LLMs). It contains approximately 10,000 JSON schemas, capturing diverse constraints and complexities.
import datasets
from datasets import load_dataset
def main():
# Inspect the available subsets of the datasetall_subsets = datasets.get_dataset_config_names("epfl-dlab/JSONSchemaBench")
print("Available subsets:"… See the full description on the dataset page: https://huggingface.co/datasets/epfl-dlab/JSONSchemaBench.sharegpt-quizz-generation-json-output
ShareGPT-Formatted Dataset for Quizz Generation in Structured JSON Output
Dataset Description
This dataset is formatted in the ShareGPT style and is designed for fine-tuning large language models (LLMs) to generate quizz in structured JSON outputs. It consists of multi-turn conversations where each response follows a predefined JSON schema, making it ideal for training models that need to produce structured data in natural language scenarios.
Usage
This dataset… See the full description on the dataset page: https://huggingface.co/datasets/Arun63/sharegpt-quizz-generation-json-output.sharegpt-structured-output-json
ShareGPT-Formatted Dataset for Structured JSON Output
Dataset Description
This dataset is formatted in the ShareGPT style and is designed for fine-tuning large language models (LLMs) to generate structured JSON outputs. It consists of multi-turn conversations where each response follows a predefined JSON schema, making it ideal for training models that need to produce structured data in natural language scenarios.
Usage
This dataset can be used to train LLMs… See the full description on the dataset page: https://huggingface.co/datasets/Arun63/sharegpt-structured-output-json.Moltbook
Moltbook AI Agent Social Media Corpus
Maintainer: Shichao Jia, Stevens Institute of Technology
Scraper source: https://github.com/jscmp4/moltbookscraper (open source)
Coverage: continuous daily collection since the platform's launch day.
This is the only Moltbook corpus spanning the platform's full history --
the February--July 2026 window cannot be collected retroactively by anyone.
Dataset Summary
A large-scale corpus of public posts, comments, and agent… See the full description on the dataset page: https://huggingface.co/datasets/jscmp4/Moltbook.resume-json-extraction-5k
Dataset Card for resume-json-extraction-5k
Dataset Description
This dataset contains 4,879 resume examples formatted for fine-tuning language models to extract structured JSON information from resume text.
Dataset Summary
The dataset consists of resume text paired with structured JSON outputs containing:
Job titles (current and previous)
Companies (current and previous)
Years of experience
Seniority level
Primary domain and industries
Core and secondary skills… See the full description on the dataset page: https://huggingface.co/datasets/sandeeppanem/resume-json-extraction-5k.json_data_extraction
Diverse Restricted JSON Data Extraction
Curated by: The paraloq analytics team.
Uses
Benchmark restricted JSON data extraction (text + JSON schema -> JSON instance)
Fine-Tune data extraction model (text + JSON schema -> JSON instance)
Fine-Tune JSON schema Retrieval model (text -> retriever -> most adequate JSON schema)
Out-of-Scope Use
Intended for research purposes only.
Dataset Structure
The data comes with the following fields:
title: The… See the full description on the dataset page: https://huggingface.co/datasets/paraloq/json_data_extraction.js_cwe_GRPO
js_cwe_GRPO
VeRL/GRPO-ready JavaScript security coding dataset generated by the simple_gen pipeline.
Each row is a harness-validated task with Node harness security/functionality tests, oracle candidate_js, and authoring guidelines (high_level_guidelines, implementational).
Files
File
Rows
Description
js_cwe_GRPO.parquet
4956
Full dataset (shuffled)
js_cwe_GRPO_train.parquet
4461
90% train split
js_cwe_GRPO_val.parquet
495
10% validation split… See the full description on the dataset page: https://huggingface.co/datasets/AetherPrior/js_cwe_GRPO.svgen_500k_rasterized_jsonified_uuided
SVGEN RJU - SVGEN 500k: Rasterized, JSONified, UUID'ed
I have selected every svg image from svgen that would rasterize under cairosvg, which is significantly less than a 1% failure rate. Under development.
Reasoning
This is the 1st of many SVG datasets I am collecting, extracting, and rasterizing in an attempt to produce a meaningfully helpful spatial reasoning and vertex manipulation model.
Usage
The rasterized images are in PNG format, as bytes. They may be… See the full description on the dataset page: https://huggingface.co/datasets/MrOvkill/svgen_500k_rasterized_jsonified_uuided.json-extraction
Rob Dixon's JSON Extraction Dataset
A synthetic dataset for training JSON extraction models, generated using Claude 3 Haiku.
Dataset Overview
This dataset contains paired examples of:
Instructions: Natural language task descriptions asking to extract information
Text documents: Source content containing information to extract
JSON outputs: Structured data extracted from the text
The dataset is designed for training smaller models on constrained context lengths, with… See the full description on the dataset page: https://huggingface.co/datasets/robdixon/json-extraction.danish-json-grpo-v1
danish-json-grpo-v1
10,015 Danish prompts for schema-directed JSON generation, built for GRPO training with a deterministic verifier (parse + key-set match + optional grounding penalty).
Task types
task_type
share
shape
extract
42%
Danish passage + schema → JSON grounded in passage
generate
26%
"Give me JSON for X with fields Y" (values open-ended)
rewrite
22%
Bullet list / semicolon-separated data → JSON with same info
fill_template
10%
JSON… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-json-grpo-v1.support-json-ru
Support-JSON-RU
Synthetic Russian SaaS support data for policy-conditioned JSON decisions and draft replies. The task supplies customer text, company policies, sourced facts and available capabilities; the model predicts a nine-field decision rather than memorizing a single company's policy.
Русский SaaS-support: обращение + правила + факты → категория, приоритет, настроение, действие, черновик ответа и эскалация.
Model · Dataset files · License
Configurations… See the full description on the dataset page: https://huggingface.co/datasets/A11Sunday/support-json-ru.reasoning-sft-JSON-structuring-and-correcting
JSON Structuring and Correcting (Reasoning SFT)
Combined dataset of 508K rows for training LLMs on structured output tasks with reasoning traces, sourced from two datasets:
Sources
tool_calling.parquet (488,461 rows)
Converted from vericava/sft-tool-calling-structured-output-v1. Multi-turn tool calling and structured output tasks including tool invocations, tool results, and final assistant responses. Includes English and Japanese content.… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/reasoning-sft-JSON-structuring-and-correcting.crawlerlm-html-to-json
CrawlerLM: HTML to JSON Extraction
A synthetic instruction-tuning dataset for training language models to extract structured JSON from HTML.
Dataset Description
This dataset contains HTML paired with structured JSON extraction tasks in chat format. It's designed for fine-tuning small language models to perform structured data extraction from messy, real-world HTML across multiple domains.
Key Features
447 examples in instruction-tuning chat format
Real HTML… See the full description on the dataset page: https://huggingface.co/datasets/espsluar/crawlerlm-html-to-json.js_tiny_codes_alpacasft_sitcom_chandlerbing_jsonl
Dataset Card for Dataset Name
This dataset card aims to be a base template for new datasets. It has been generated using this raw template.
Dataset Details
Dataset Description
Curated by: [More Information Needed]
Funded by [optional]: [More Information Needed]
Shared by [optional]: [More Information Needed]
Language(s) (NLP): [More Information Needed]
License: [More Information Needed]
Dataset Sources [optional]
Repository: [More… See the full description on the dataset page: https://huggingface.co/datasets/deepakkarkala/sft_sitcom_chandlerbing_jsonl.reasoning-sft-interstellarninja-json-mode-reasoning-160K
json-mode-reasoning (converted)
Converted version of interstellarninja/json-mode-reasoning, filtered to 20,474 rows with valid <think> reasoning traces.
Format
Each row has three columns:
input — list of dicts [{"role": "system/user", "content": "..."}, ...] (conversation turns ending on the last user turn, includes system prompt with JSON schema)
response — assistant response string with <think> reasoning block followed by JSON output
source — fixed as… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/reasoning-sft-interstellarninja-json-mode-reasoning-160K.synthetic-info-extract-json
Raw text to json object (synthetic)
Amad Zarak
February 28, 2026
Created using gpt-oss-120b on h200 sxm
Zero-shot JSON schema deduction & universal information extraction.
80,664 rows
Figured others could use this since it is basically impossible to find massive raw-text-to-structured-json datasets for training extraction engines.
About 30k of the raw outputs hit the token limit and malformed, but I ran a massive salvage sweep on the raw outputs using json-repair to force the… See the full description on the dataset page: https://huggingface.co/datasets/amadzarak/synthetic-info-extract-json.amadesus-trl-assistant-dataset-v2-0
AMADEUS_TRL_DATASET
Dataset Description
amadesu_trl_assistant_dataset is designed to train intelligent assistants in evaluating the Technology Readiness Level (TRL) in the field of agriculture, using the TRL metric developed by NASA. The dataset is organized into two parts:
Conceptual Knowledge Dataset: Provides essential knowledge about TRL concepts and definitions, levels, objectives, and goals for each level, as well as related technological development activities.… See the full description on the dataset page: https://huggingface.co/datasets/JsBetancourt/amadesus-trl-assistant-dataset-v2-0.math500-deepseek-r1-distill-qwen-1.5b
Dataset Card for "math500-deepseek-r1-distill-qwen-1.5b"
Dataset Summary
This dataset is a distilled version of the MATH500 dataset, augmented with reasoning-based responses generated by the deepseek-r1-distill-qwen-1.5b language model. The dataset is designed to evaluate and improve the mathematical reasoning capabilities of LLMs through step-by-step solutions and final answers.
Each example consists of:
The original problem statement from MATH500
The reference solution… See the full description on the dataset page: https://huggingface.co/datasets/jsm0424/math500-deepseek-r1-distill-qwen-1.5b.geo-perturbation-grouping-test
GEO Gene-Perturbation Grouping — Held-out Test
The held-out evaluation set for the GEO gene-perturbation grouping task. The training
and validation data live in the companion private repo
jsoul/geo-perturbation-grouping-train,
which is where the full task description, rulebook, and data-format documentation are
maintained.
⚠️ Intended as a clean evaluation benchmark — avoid using it for training or model
selection.
Contents
300 examples, one test split.
300… See the full description on the dataset page: https://huggingface.co/datasets/jsoul/geo-perturbation-grouping-test.math500-deepseek-r1-distill-qwen-14b
Dataset Card for "math500-deepseek-r1-distill-qwen-14b"
Dataset Summary
This dataset is a distilled version of the MATH500 dataset, augmented with reasoning-based responses generated by the deepseek-r1-distill-qwen-14b language model. The dataset is designed to evaluate and improve the mathematical reasoning capabilities of LLMs through step-by-step solutions and final answers.
Each example consists of:
The original problem statement from MATH500
The reference solution… See the full description on the dataset page: https://huggingface.co/datasets/jsm0424/math500-deepseek-r1-distill-qwen-14b.geo-perturbation-grouping-train
GEO Gene-Perturbation Grouping — Train / Validation
Supervised fine-tuning (SFT) data for teaching an LLM to read the sample metadata of a
NCBI GEO gene-expression series and organize its
samples into valid genetic-perturbation case/control experimental groups.
This repository holds the train and validation splits. The held-out test set
lives in a separate private repo: jsoul/geo-perturbation-grouping-test.
What the task is
Given a GEO study (its… See the full description on the dataset page: https://huggingface.co/datasets/jsoul/geo-perturbation-grouping-train.deepscaler-preview-kogutenberg-multilingualgeo-perturbation-grouping-production
GEO Gene-Perturbation Grouping — Production / Inference Inputs
The production inference corpus for the GEO gene-perturbation grouping task: a large set
of NCBI GEO studies formatted as model prompts, to be
run through a fine-tuned model (e.g.
jsoul/geo-perturbation-grouping-qwen3-30b-a3b
or jsoul/geo-perturbation-grouping-qwen3-4b)
to generate case/control perturbation groupings at scale.
⚠️ Unlabeled — inputs only. Every row contains just a system + user message and
no… See the full description on the dataset page: https://huggingface.co/datasets/jsoul/geo-perturbation-grouping-production.experimental-paper-json-xtractionjs_repl_ai_alpacaexperimental-paper-json-xtraction-2Thinker-JSONUse for whatever you want.
Made to replicate the thought traces of OpenAI's o1, I'll release RL datasets including DPO soon enough.
For fine-tuning smaller models such as Google's google/gemma-2-2b-it with this dataset, I recommend fine-tuning for 2-3 epochs, the loss will be at around 1.6 at the beginning, and 1.3 by the end of the training job with learning rate of 2e-6.
Suggested system prompt:
Always respond in strict JSON format with a reasoning_steps array and a response field. Each… See the full description on the dataset page: https://huggingface.co/datasets/minchyeom/Thinker-JSON.job_text_to_json_llama-2