datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
vllm-control-arena
vLLM Main Tasks Dataset
AI coding tasks generated from vLLM git commits
Dataset Description
This dataset contains 6801 coding tasks automatically generated from git commits in the vLLM repository. Each task represents a real-world coding challenge derived from actual development work.
Dataset Structure
The dataset contains the following columns:
commit_hash: The git commit hash
parent_hash: The parent commit hash
commit_title: The original commit… See the full description on the dataset page: https://huggingface.co/datasets/RoganInglis/vllm-control-arena.MMFineReason-Full-2.3M-Qwen3-VL-235B-Thinking
MMFineReason-Full-2.3M
The Complete Pre-Selection Dataset — Before Quality Filtering
📖 Overview
MMFineReason-Full-2.3M is the complete pre-selection dataset containing 2.3M samples and 8.8B solution tokens, generated through our reasoning distillation pipeline before the data selection stage. This dataset includes all samples that passed basic template and length validation, but have not undergone correctness verification filtering.
🎯 Key Characteristics… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-Full-2.3M-Qwen3-VL-235B-Thinking.MMFineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-1.8M-Qwen3-VL-235B-Thinking.MMFineReason-Full-2.3M-Qwen3-VL-235B-Thinking
MMFineReason-Full-2.3M
The Complete Pre-Selection Dataset — Before Quality Filtering
📖 Overview
MMFineReason-Full-2.3M is the complete pre-selection dataset containing 2.3M samples and 8.8B solution tokens, generated through our reasoning distillation pipeline before the data selection stage. This dataset includes all samples that passed basic template and length validation, but have not undergone correctness verification filtering.
🎯 Key Characteristics… See the full description on the dataset page: https://huggingface.co/datasets/ericktwo/MMFineReason-Full-2.3M-Qwen3-VL-235B-Thinking.FineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/NarsAI/FineReason-1.8M-Qwen3-VL-235B-Thinking.sat-vl-sft-training-ready-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-vl-sft-training-ready-v1.MMFineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/Sandeepthakur/MMFineReason-1.8M-Qwen3-VL-235B-Thinking.vi-dataset-for-pretrain
Dataset Card for "vi-dataset-for-pretrain"
This is a combination of multiple Vietnamese dataset for pretraining CLMs such as GPT, GPT2, etc.
The dataset consists of:
vietgpt/covid_19_news_vi
hieunguyen1053/binhvq-news-corpus
oscar (unshuffled_deduplicated_vi)
vietgpt/wikipedia_vi
Dataset info
Splits
N.o examples
Size
Train
23,891,116
77.36 GB
Validation
1,257,428
4.06 GB
Total
25,148,544
81.43 GB
pretrain-dataset-raw-10M
Pretrain Dataset (Text)
This dataset contains preprocessed text documents ready for LLM pretraining.
Dataset Details
Property
Value
Documents
10,000,000
Processed
10000000
Shards
21
Created
2025-12-09
Dataset Structure
Each sample contains:
text: The document text
source: Source dataset identifier
id: Unique document ID
Usage
from datasets import load_dataset
dataset = load_dataset("tvu-vlinhd11/pretrain-dataset-raw-10M")… See the full description on the dataset page: https://huggingface.co/datasets/tvu-vlinhd11/pretrain-dataset-raw-10M.gemini_public_mmr1
PRISM Public SFT Data
Overview
PRISM Public SFT Data is the public supervised fine-tuning data collection used in the PRISM project.PRISM studies the distributional drift problem in the standard SFT → RLVR post-training pipeline for large multimodal models. Before the distribution alignment and RLVR stages, we first use large-scale public multimodal demonstrations to obtain a broad SFT initialization.
This dataset serves as the public SFT data source for the… See the full description on the dataset page: https://huggingface.co/datasets/prism-vlm/gemini_public_mmr1.sat-vl-sft-postprocessed-merged-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-vl-sft-postprocessed-merged-v1.loong_c4A filtered subset of C4-en containing 3,584,358 pages that are at least 16,000 characters long, useful for training models with longer context windows.
Arabic-VLM-Full-Pearl
💎 The Arabic VLM Dataset (Full Pearl Edition)
This repository contains the full, unreviewed dataset comprising 309K multimodal examples. This data was generated automatically using the agentic pipeline developed for the Pearl project, as described in our paper.
Disclaimer: This is the raw, synthetic data that has not been subject to human review. It was generated as part of the data creation process and is released for research purposes. It may contain noise, errors, or… See the full description on the dataset page: https://huggingface.co/datasets/MohamedRashad/Arabic-VLM-Full-Pearl.MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking
MMFineReason-SFT-123K
The Hardest 7% — Less Data, More Reasoning
📖 Overview
MMFineReason-SFT-123K is a difficulty-filtered subset of MMFineReason-1.8M, containing only the hardest 7% of samples where Qwen3-VL-4B-Thinking consistently fails (pass rate = 0).
🎯 Key Highlights
123K Challenging Samples: Only instances where a 4B thinking model fails all 4 inference attemptsEfficient Training: Comparable performance to full 1.8M dataset with only 7% of… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking.long_c4A filtered subset of C4-en containing 13,688,429 pages that are at least 8,000 characters long, useful for training models with longer context windows.
MMFineReason-1.8M-Qwen3-VL-235B-Thinking
MMFineReason
Closing the Multimodal Reasoning Gap via Open Data-Centric Methods
Average score across mathematical reasoning and multimodal understanding benchmarks.
📖 Overview
MMFineReason is a large-scale, high-quality multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring detailed reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking.
🎯 Key Highlights
1.8M High-Quality Samples with 5.1B Solution Tokens… See the full description on the dataset page: https://huggingface.co/datasets/dans25275/MMFineReason-1.8M-Qwen3-VL-235B-Thinking.UTS_VLC
Dataset Card for Vietnamese Legal Corpus (UTS_VLC)
A curated corpus of Vietnamese Laws and Codes (Luật, Bộ luật) and the Constitution,
maintained by Underthesea NLP. The flagship 2026 split is a
verified in-force snapshot — every document is currently in force, de-duplicated, and validated
against Vietnam's official legal database vbpl.vn.
Dataset Details
Dataset Description
UTS_VLC contains the full text of Vietnamese legislation at the top of the… See the full description on the dataset page: https://huggingface.co/datasets/undertheseanlp/UTS_VLC.rl_dataset
PRISM RL Dataset
Overview
PRISM RL Dataset contains the training data used for the PRISM alignment and RLVR stages.
PRISM studies the distributional drift problem in the standard SFT → RLVR post-training pipeline for large multimodal models. Instead of directly applying RLVR after SFT, PRISM inserts an intermediate Distribution Alignment / Pre-alignment stage based on black-box on-policy distillation.
The overall pipeline is:
SFT → PRISM Alignment → RLVR
This… See the full description on the dataset page: https://huggingface.co/datasets/prism-vlm/rl_dataset.Leesplank-vloeiend-nl-curriculum-cp2
Leesplank NL — Embeddings & Clusters (Checkpoint 2)
5.39 million Dutch texts, each annotated with a 384-dimensional semantic embedding,
a K-means cluster assignment, and — for a stratified sample of ~100,800 rows — a
measured BF16 perplexity score from IBM Granite 4.0 Micro (3B Dense).
This checkpoint is the analytical core of a project with two concrete goals: running
high-quality Dutch language processing on consumer hardware, and doing it in a way
that is transparent enough for… See the full description on the dataset page: https://huggingface.co/datasets/MichielBuisman/Leesplank-vloeiend-nl-curriculum-cp2.gemini_distill
PRISM Gemini Distill
Overview
PRISM Gemini Distill is our self-distilled multimodal reasoning dataset collected from Gemini 3 Flash for the PRISM project.
PRISM studies the distributional drift problem in the standard SFT → RLVR post-training pipeline. To mitigate this issue, PRISM introduces an intermediate Distribution Alignment / Pre-alignment stage before RLVR:
SFT → Distribution Alignment / Pre-alignment → RLVR
This dataset provides high-quality Gemini 3… See the full description on the dataset page: https://huggingface.co/datasets/prism-vlm/gemini_distill.looong_c4A filtered subset of C4-en containing 835,400 pages that are at least 32,000 characters long, useful for training models with longer context windows.
guitar_tabDataset of music tablature, in alphaTex (https://alphatab.net/docs/alphatex)
format, converted from Guitar Pro files (gp3, gp4, gp5, which are downloaded
from https://rutracker.org/forum/viewtopic.php?t=2888130MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking-QR-max4096
Derived dataset note
This dataset was derived from OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking as a part of arxiv.org/abs/2603.22276.
Field changes:
question -> query
qwen3vl_235b_thinking_response -> response
image -> images (single-item list)
added tok_len, computed with tokenizer Qwen/Qwen3-8B on query + '\n\n' + response
add_special_tokens=False
The original README content is preserved below.
MMFineReason-SFT-123K
The Hardest 7% — Less Data, More Reasoning… See the full description on the dataset page: https://huggingface.co/datasets/eyes-ml/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking-QR-max4096.tsiolkovsky-papers
Tsiolkovsky Papers: the complete personal archive as a machine-readable corpus
Machine transcriptions of all 51,008 sheets of fond 555 of the Archive of the
Russian Academy of Sciences — the personal archive of Konstantin Tsiolkovsky
(1857–1935), who derived the rocket equation and described the multistage rocket
decades before anyone could test either.
The archive had been scanned and put online, but without a catalogue you could
query, full-text search, or a dataset. This is… See the full description on the dataset page: https://huggingface.co/datasets/vladimirbesk/tsiolkovsky-papers.re-tutor-protection-mechanisms
RE-Tutor: Protection-Mechanism Analysis Dataset
Instruction-tuning dataset teaching a model to analyze protection mechanisms
(anti-debug, anti-VM, anti-tamper, anti-dump, obfuscation, timing) from code
evidence and emit structured expert analysis.
Schema
Each sample pairs input (code evidence) with output (structured analysis):
input.code_snippet: C source, decompiler-style pseudocode, or x86/x64 assembly
input.imports_pool: mixed DLL!API imports (includes… See the full description on the dataset page: https://huggingface.co/datasets/vluxblaring/re-tutor-protection-mechanisms.VLM-SFTvietnamese-sft-10k
Vietnamese Instruction-Following Dataset (10K)
This dataset comprises 10,000 Vietnamese instruction-style prompt-response pairs curated for supervised fine-tuning (SFT) of language models. It aims to improve conversational and instruction-following abilities in the Vietnamese language, with coverage across diverse social, cultural, and emotional contexts.
Format: JSONL (one object per line)
Fields: "prompt" (instruction or user message), "response" (assistant reply)
Language:… See the full description on the dataset page: https://huggingface.co/datasets/vlinhd11/vietnamese-sft-10k.CIK-Bench
CIK-Bench
Official dataset for Your Agent, Their Asset: A Real-World Safety Analysis
of OpenClaw.
CIK-Bench evaluates OpenClaw — the most widely deployed personal AI
agent in early 2026 — against persistent-state poisoning attacks. It
implements the CIK taxonomy, a unified framework that organizes
OpenClaw's persistent state into three dimensions:
Capability — executable skills (SKILL.md, .sh, .py)
Identity — persona, values, and behavioral configuration
(SOUL.md, IDENTITY.md… See the full description on the dataset page: https://huggingface.co/datasets/UCSC-VLAA/CIK-Bench.deepscaler-teacher-sft-vllm-official-40k
DeepScaleR teacher SFT vLLM official 40k
Generated run: exp_003_vllm_official_brainlab_2gpu.
Summary
{
"num_examples": 40300,
"sft_dir": "data/processed/deepscaler/teacher_sft/exp_003_vllm_official_brainlab_2gpu",
"parse_rate": 0.9999751861042183,
"correct_rate": 0.5728039702233251,
"format_rate": 0.005955334987593052,
"mean_reward": 0.42432258064534184,
"deepscaler_mean_reward": 0.6266997518610422,
"deepscaler_match_mean_reward":… See the full description on the dataset page: https://huggingface.co/datasets/ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k.medical_sft_crawl_vi_10k_v1
ViMed-SFT: Vietnamese Medical Conversational Dataset
Dataset Description
ViMed-SFT is a high-quality Vietnamese medical conversational dataset designed for Supervised Fine-Tuning (SFT) of Large Language Models. The dataset contains medical Q&A conversations between users and a virtual medical assistant.
Dataset Summary
Attribute
Value
Language
Vietnamese
Domain
Healthcare / Medical
Task
Conversational AI, Instruction Tuning
Samples… See the full description on the dataset page: https://huggingface.co/datasets/vlinhd11/medical_sft_crawl_vi_10k_v1.