datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
lingshu_training_data_medical_domain
Website
🤖 7B Model
🤖 8B Model based on InternVL3
🤖 32B Model
MedEvalKit
Technical Report
Lingshu MCP
Lingshu Medical MLLM Training Data (Medical Domain)
This dataset contains the medical-domain training data used in the multi-stage training of the Lingshu Medical Multimodal Large Language Model (MLLM). General-domain data has been removed; only medical data is included.
The training… See the full description on the dataset page: https://huggingface.co/datasets/lingshu-medical-mllm/lingshu_training_data_medical_domain.DR3-EvalDR3-Eval: Towards Realistic and ReproducibleDeep Research Evaluation
✨ Overview
DR³-Eval is a realistic, reproducible, and multimodal evaluation benchmark for Deep Research Agents, focusing on multi-file report generation tasks.
Existing benchmarks face a fundamental tension between realism, controllability, and reproducibility when evaluating deep research agents. DR³-Eval addresses this through the following design:… See the full description on the dataset page: https://huggingface.co/datasets/NJU-LINK/DR3-Eval.WebCompass
WebCompass
A unified multimodal benchmark for evaluating LLMs' ability to generate, edit, and repair functional web pages. WebCompass spans three input modalities — text design documents, reference screenshots, and video demonstrations — and three task families — generation, editing, and repair.
GitHub: NJU-LINK/WebCompass
Project Page: nju-link.github.io/WebCompass
Quick Start
from datasets import load_dataset
# Generation tasks (existing)
ds_text =… See the full description on the dataset page: https://huggingface.co/datasets/NJU-LINK/WebCompass.linux-command-dataset
Linux Command Dataset
A comprehensive dataset of Linux command examples designed for training language models. The dataset pairs natural language descriptions with their corresponding shell commands, covering a wide range of common operations. This dataset was trained on Llama 3.2 1b, and the final version has been uploaded to Hugging Face: mecha-org/linux-command-generator-llama3.2-1b.
Dataset Statistics
This table reflects the actual number of command examples in… See the full description on the dataset page: https://huggingface.co/datasets/mecha-org/linux-command-dataset.LinguaNovaReasonMed
ReasonMed: A 370K Multi-Agent Generated Dataset for Advancing Medical Reasoning
📄 Paper |
💻 Code |
📊 Dataset
ReasonMed is the largest open-source medical reasoning dataset to date, containing 370 K high-quality question–answer examples with multi-step chain-of-thought (CoT) rationales and concise summaries. We distilled these from 1.75 M initial reasoning paths generated by three competitive large-language models (Qwen-2.5-72B, DeepSeek-R1-Distill-Llama-70B, and… See the full description on the dataset page: https://huggingface.co/datasets/lingshu-medical-mllm/ReasonMed.IndicTalk
IndicTalk:Code-Mixed Conversational Persona-Based Dataset
This dataset contains multi-turn, persona-driven, code-mixed conversations
generated from real news articles, across 9 Indian
languages, in two script variants:
Native — conversations written in the language's native script,
code-mixed with Romanized English words.
Romanized — conversations fully Romanized (Latin script), code-mixed
with English.
Each language has its own config, loadable independently, e.g.:
from… See the full description on the dataset page: https://huggingface.co/datasets/LingoIITGN/IndicTalk.Linux-terminal-tool-calling
Linux Terminal Tool Calling Dataset (Linux-terminal-tool-calling)
This dataset is designed for training and fine-tuning AI agents on tool calling, reasoning, and command execution specifically for standard Linux terminal utilities and system administration tasks. It transforms raw Linux terminal command records into a structured multi-turn conversation format featuring detailed chain-of-thought/reasoning content and OpenAI/OpenClaw-style function calling.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/iselabvn/Linux-terminal-tool-calling.open-perfectblend-qwen3-4b-nonthinking
Open PerfectBlend Qwen3-4B Non-Thinking
This dataset contains 1,349,812 training conversations derived from mlabonne/open-perfectblend. Each assistant turn was regenerated sequentially with Qwen/Qwen3-4B, conditioned on the preceding conversation, with thinking disabled.
Data preparation
Empty or otherwise invalid source conversations were removed before a deterministic train/evaluation split. The split used seed 42 and a held-out fraction of 0.05. Only the 1,349… See the full description on the dataset page: https://huggingface.co/datasets/linYD0718/open-perfectblend-qwen3-4b-nonthinking.TexAes
Dataset Card for TexAes
TexAes is the first aesthetic dataset in the LLM domain, containing a total of 50,390 prompts. It is curated by an aesthetic data generation pipeline leveraging GPT-4o for aesthetic polishing, as described in our paper "Textual Aesthetics in Large Language Models."
To address the challenge of generating high-quality aesthetic preference data, we developed a scalable aesthetic data generation pipeline. This pipeline utilizes GPT-4o to enhance the aesthetic… See the full description on the dataset page: https://huggingface.co/datasets/lingjie23/TexAes.LogiTraj-Benchmark
LogiTraj Benchmark
Dual license. Dataset material is licensed under CC BY 4.0;
software in evaluation/code/ is licensed under Apache-2.0. See LICENSE,
LICENSE-DATA, and LICENSE-CODE. Commercial-model raw outputs are not
included.
Synthetic Chinese enterprise logistics sandboxes, tasks, documents, versioned evaluators, verdicts, and Core/Silver/Audit quality views.
Task-family coverage is source-faithful rather than imputed: the 20260628_v45, 20260628_v46, and 20260628_v50 SFT… See the full description on the dataset page: https://huggingface.co/datasets/Linmumu009/LogiTraj-Benchmark.open-perfectblend-kimi-linear-regen
Open PerfectBlend Kimi Linear Regen
This dataset regenerates the assistant messages in
mlabonne/open-perfectblend
with Kimi-Linear-48B-A3B-Instruct. It is intended for speculative-decoding
drafter training and related research.
Generation
Source conversation structure and user messages: mlabonne/open-perfectblend
Target model: Kimi-Linear-48B-A3B-Instruct
Temperature: 0.7
Maximum new tokens per assistant turn: 8192
Assistant turns were regenerated sequentially.… See the full description on the dataset page: https://huggingface.co/datasets/Dogacel/open-perfectblend-kimi-linear-regen.cve-single-line-fixes
Dataset Card for Dataset Name
Dataset Summary
This dataset card aims to be a base template for new datasets. It has been generated using this raw template.
Supported Tasks and Leaderboards
[More Information Needed]
Languages
[More Information Needed]
Dataset Structure
Data Instances
[More Information Needed]
Data Fields
[More Information Needed]
Data Splits
[More Information Needed]
Dataset Creation… See the full description on the dataset page: https://huggingface.co/datasets/lambdasec/cve-single-line-fixes.Alpsbench🚨 We invite everyone to checkout our AlpsBench on 🤗HuggingFace, focusing on real-world dialogue memorization and implicit user preferences!
This is the official Huggingface repository of the paper AlpsBench: An LLM Personalization Benchmark for Real-Dialogue Memorization and Preference Alignment and the AlpsBench benchmark.
We present AlpsBench, a new LLM personalization benchmark derived from real-world human–LLM dialogues. While existing benchmarks rely heavily on synthetic dialogues that… See the full description on the dataset page: https://huggingface.co/datasets/linston666/Alpsbench.ProBench
ProBench
A 4-choice multiple-choice benchmark of 529 expert-level questions across five professional domains, sourced entirely from Stack Exchange communities (CC-BY-SA 4.0). Designed to evaluate LLMs where standard benchmarks (MMLU, HumanEval, HellaSwag) are now saturated above 90%.
Why ProBench?
Benchmark
Frontier model score
Status
MMLU
>90%
Saturated
HellaSwag
>95%
Saturated
HumanEval
>90%
Saturated
ProBench
TBD
Active
Questions are… See the full description on the dataset page: https://huggingface.co/datasets/lin99/ProBench.ChiEngMixBench-Dataset
ChiEngMixBench v0.2.0
Paper: arXiv:2601.16217Code and frozen release: GitHubGitHub release: v0.2.0
ChiEngMixBench evaluates terminology-form choice in Chinese AI/CS discourse. It contains controlled Chinese-English minimal pairs, item-level model outputs, anonymized human ratings, and auditable analysis code.
This release deliberately separates two views:
Paired terminology choice: whether an open-weight model assigns higher length-normalized sequence likelihood to an English… See the full description on the dataset page: https://huggingface.co/datasets/AI-Ling00/ChiEngMixBench-Dataset.connect4
Dataset Card for Connect4 Reasoning Task
1. Dataset Summary
This dataset is dynamically constructed using the GAMEBoT framework in the Connect4 game. It is designed to evaluate Large Language Models' (LLMs) ability in symbolic reasoning, board state parsing, and lookahead planning.
By serializing 6 × 7 grid states into text-based formats, this dataset challenges models to identify winning topologies in a deterministic environment with a state-space complexity of… See the full description on the dataset page: https://huggingface.co/datasets/linius/connect4.linux-commandsLogiTraj-Trajectories
LogiTraj Trajectories
License: This repository's dataset material is licensed under
cc-by-4.0. See LICENSE. Commercial-model raw outputs are not
included.
Sanitized open-model agent trajectories (57,514 records); hidden reasoning, provider request metadata, exact timestamps, and internal identifiers are removed.
The public configurations contain 28,500 complete converted SFT conversations and 29,014 complete test observable-event rollouts.
Contact-shaped values are replaced… See the full description on the dataset page: https://huggingface.co/datasets/Linmumu009/LogiTraj-Trajectories.linkedin-natural-150
LinkedIn Natural 150 — gemma-3-1b-it style tuning
150 curated LinkedIn posts in a simple, conversational, twitter-like style (no cringe drama).
Composition: 44 twitter-gold short (<50w) + 84 core short-medium (50-80w) + 20 medium-long (80-150w) + 2 long gold. Avg 60.7w.
Format per line (JSONL):
{"prompt": "Write a LinkedIn post about: <topic>", "response": "<natural post>"}
Files:
data/final_dataset.jsonl — 150 full examples (use this)
data/train.jsonl — 138 train… See the full description on the dataset page: https://huggingface.co/datasets/harsh-jos/linkedin-natural-150.mathematics_competition
Mathematics Competition Evaluation
Competition-level mathematics evaluation dataset with 3-run predictions from Gemini model.
Dataset Structure
Each row contains:
uuid: unique identifier
question: math competition problem
answer: ground truth answer
source: problem source
run_0, run_1, run_2: each a dict with:
prediction: model's answer
stream_output: list of stream output segments
stream_output_kinds: list of output kinds (thought/text/tool_call)
correct: whether… See the full description on the dataset page: https://huggingface.co/datasets/LinhIcey/mathematics_competition.linghuchong支持ChatHaruhi2 的令狐冲数据，可以使用如下方式调用
from chatharuhi import ChatHaruhi
chatbot = ChatHaruhi( role_from_hf = 'chengli-thu/linghuchong', \
llm = 'openai')
response = chatbot.chat(role='小师妹', text = '冲哥。')
print(response)
上传者: 李鲁鲁
更具体的信息，见 ChatHaruhi
欢迎加入我们的 众筹角色创建项目
Citation引用
Please cite the repo if you use the data or code in this repo.
@misc{li2023chatharuhi,
title={ChatHaruhi: Reviving Anime Character in Reality via Large Language Model}… See the full description on the dataset page: https://huggingface.co/datasets/chengli-thu/linghuchong.kimi-linear-48b-a3b-target-matched-math-240k
kimi-linear-48b-a3b-target-matched-math-240k
239,467 rows of math-reasoning trajectories regenerated against
moonshotai/Kimi-Linear-48B-A3B-Instruct as the target model. Used to train DFlash
speculative-decoding drafters in
la-draftery.
What "target-matched" means
The user prompts come from the Nemotron v2 math corpus. The assistant
completions in this dataset are the target model's own outputs — each
prompt was sent to moonshotai/Kimi-Linear-48B-A3B-Instruct and its… See the full description on the dataset page: https://huggingface.co/datasets/Moonlight556/kimi-linear-48b-a3b-target-matched-math-240k.linear-bench-mini
Agent-Diff: Linear Bench Mini
This dataset is part of the Agent-Diff benchmark, presented in the paper Agent-Diff: Benchmarking LLM Agents on Enterprise API Tasks via Code Execution with State-Diff-Based Evaluation.
Website | GitHub | Paper
Context
The Linear Bench suite runs inside the Agent Diff isolation engine, with its own Postgres schema replaying the Linear GraphQL API. Agents interact via Linear's public surface area to satisfy CRUD-style tasks (create issues… See the full description on the dataset page: https://huggingface.co/datasets/hubertmarek/linear-bench-mini.Ling-v2.6-Flash-Distilled-15K
⚡ Ling-2.6-Flash-Distilled-15K
Dataset Summary
Ling-2.6-Flash-Distilled-15K is a supervised fine-tuning dataset for logic-oriented distillation. The prompts to the questions come from Jackrong/GLM-5.1-Reasoning-1M-Cleaned, and the answers were generated using the only Ling-2.6-Flash teaching model.
Dataset Details
Dataset
constructai/Ling-2.6-Flash-Distilled-15K
Source questions
Jackrong/GLM-5.1-Reasoning-1M-Cleaned
Teacher model… See the full description on the dataset page: https://huggingface.co/datasets/constructai/Ling-v2.6-Flash-Distilled-15K.ai-paper-intellectual-lineage-2023
Intellectual Lineage of Impactful AI Research Papers (2023-2024)
Dataset Description
This dataset contains 20 impactful AI research papers published between 2022-2024, along with their intellectual lineage - tracing 1-2 key prior works each paper builds upon, and a ~300-word paragraph explaining the relationship between the current work and its foundations.
Purpose
Understanding how research ideas evolve and build upon prior work is crucial for:
Researchers… See the full description on the dataset page: https://huggingface.co/datasets/AmberLJC/ai-paper-intellectual-lineage-2023.Linalg-Spec-30
Linalg-Spec-30
Accepted to NeurIPS 2026 (Evaluations & Datasets Track)Paper (arXiv) · Code (GitHub) · All six datasets (collection)
Hand-authored NL→MLIR pairs for linalg named ops under memref semantics (n=30).
This dataset is one of six NL→MLIR benchmarks released with the NeurIPS 2026 Evaluations & Datasets Track paper Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR (arXiv:2607.18254). The full suite… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/Linalg-Spec-30.cross_lingual_transfer_dialog_generationcross-lingual transfer in dialog generation
Chinese dialogs in movie domain: Chinese_corpus/train.jsonl, Chinese_corpus/dev.jsonl, Chinese_corpus/test.jsonl. The sizes are 500/50/500.
English dialogs in movie domain: English_corpus/train.jsonl, English_corpus/dev.jsonl. The sizes are 400k/20k.
Chinese dialogs for test in music/book/tech domain: other_domains/music.test.jsonl, other_domains/book.test.jsonl, other_domains/tech.test.jsonl. The sizes are 500/500/500.
Citation… See the full description on the dataset page: https://huggingface.co/datasets/lorashen/cross_lingual_transfer_dialog_generation.CV-LinkedIn
Sample Dataset Structure
{
"image_source": "local-files:///curriculum_vitae.png",
"profile_data": {
"NAME": {
"text": "Muhammad Fawwaz Humam",
"bbox": [10.5, 5.2, 35.0, 4.1]
},
"DOMISILI": {
"text": "Bekasi, Indonesia",
"bbox": [10.5, 9.3, 20.1, 2.5]
},
"SUMMARY": {
"text": "A passionate Full-stack Web Developer and Machine Learning Engineer experienced in building… See the full description on the dataset page: https://huggingface.co/datasets/mfhumam/CV-LinkedIn.linhhuonglinux-office-dataset-v3
🚀 Linh Hương Linux Office Dataset V3 (Master/Production Ready)
Đây là bộ dữ liệu khổng lồ thế hệ mới nhất dành riêng cho việc huấn luyện Linh Hương Linux AI.
Tập dữ liệu này được chia làm 2 cấu hình (Configs) riêng biệt để tránh lỗi xung đột cấu trúc (Schema Conflict):
1. Cấu hình SFT (sft)
Chứa 1,783 tình huống Supervised Fine-Tuning, làm sạch 100% bằng Heuristic Rejection Sampling.
Multi-turn Chat: Hội thoại nhiều lượt (nhớ bối cảnh).
Function Calling: Gọi hàm điều… See the full description on the dataset page: https://huggingface.co/datasets/linhhuonglinux/linhhuonglinux-office-dataset-v3.