datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Multilingual-Thinking
Dataset summary
Multilingual-Thinking is a reasoning dataset where the chain-of-thought has been translated from English into one of 4 languages: Spanish, French, Italian, and German. The dataset was created by sampling 1k training samples from the SystemChat subset of SmolTalk2 and translating the reasoning traces with another language model.
This dataset was used in the OpenAI Cookbook to fine-tune the OpenAI gpt-oss models.
You can load the dataset using:
from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/Multilingual-Thinking.thinking_fmb_dataset_lerobot_output_qwen3vlMMFineReason-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.thinking_droid_lerobot_output_qwen3vlPP
SteelBench: A Diagnostic Benchmark for Vision-Language Models in Industrial Safety Monitoring
SteelBench is a diagnostic benchmark of densely annotated CCTV clips from an
operating integrated steel plant. It is designed to evaluate vision-language
models (VLMs) on real-world industrial action recognition, PPE assessment,
and safety-violation detection — under naturally occurring degradation
(dust, glare, steam, low light), at distances and crowdedness levels that
curated… See the full description on the dataset page: https://huggingface.co/datasets/ThinkingHub/PP.llm-jp-4-thinking-sft-data
llm-jp-4-thinking-sft-data
Overview
This dataset is a supervised fine-tuning (SFT) dataset used to train llm-jp-4-*-thinking models.
This dataset is constructed by extracting prompts from multiple data sources and generating reasoning processes and final responses using gpt-oss-120b.
The splits reasoning_low, reasoning_medium, and reasoning_high correspond to different reasoning effort settings used during generation with gpt-oss-120b.
To support the continued development… See the full description on the dataset page: https://huggingface.co/datasets/llm-jp/llm-jp-4-thinking-sft-data.explore-thinking-models-internalCodeX-2M-Thinking
Modotte
Note: This dataset is part of the lineup CodeX by Modotte. You can get lots of datasets in this same lineup, with the main focus on providing very high-quality datasets for model training and fine-tuning.
This dataset is fully synthetic, curated from high-quality public sources and enhanced with synthetic data generated using both closed and open-source models. It serves as a strong foundation for instruction-based model tuning and fine-tuning, offering one of the… See the full description on the dataset page: https://huggingface.co/datasets/Modotte/CodeX-2M-Thinking.ioi-eval-openrouter_anthropic_claude-3_7-sonnet_thinking-prompt-mem-limitioi-eval-openrouter_google_gemini-2_0-flash-thinking-exp-prompt-mem-limitthinking_furniture_bench_dataset_lerobot_output_qwen3vlMMFineReason-SFT-586K-Qwen3-VL-235B-Thinking
MMFineReason-SFT-586K
The Hardest 33% — Less Data, More Reasoning
📖 Overview
MMFineReason-SFT-586K is a difficulty-filtered subset of MMFineReason-1.8M, containing the hardest 33% of samples where Qwen3-VL-4B-Thinking do not consistently succeed. (pass rate ≠ 1).
Specifically, this subset removes all easy samples (pass rate = 1) under Qwen3-VL-4B-Thinking, retaining only instances that require non-trivial multimodal reasoning.
🎯 Key Highlights
586K… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-SFT-586K-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.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.ThinkingBox-Bench
ThinkingBox-Bench
ThinkingBox-Bench is an executable benchmark for evaluating whether tool-using
LLM agents can reliably complete stateful business workflows. Version 1.0
contains 507 tool-agent-user tasks across retail and e-commerce, travel and
hospitality, auto insurance, neobank support, and consulting IT/HR support.
This dataset repository provides a browsable representation of the benchmark.
The executable benchmark, tool servers, and supporting fixtures are maintained
in… See the full description on the dataset page: https://huggingface.co/datasets/microsoft/ThinkingBox-Bench.VLAA-Thinking
SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models
🌐 Project Page
• 📄 Arxiv
• 💻 Code
🤗 VLAA-Thinker Family
• 🤔 VLAA-Thinking Dataset
🤗 VLAA-Thinker-Qwen2.5-3B
• 🤗 VLAA-Thinker-Qwen2.5-7B
Both VLAA-Thinker-Qwen2.5-3B and VLAA-Thinker-Qwen2.5-7Bachieve SOTA performance on OpenCompass Multimodal Reasoning Leaderboard as of April 7th, 2025.
Contents
Quick Start 🚀… See the full description on the dataset page: https://huggingface.co/datasets/UCSC-VLAA/VLAA-Thinking.thinking_stanford_hydra_dataset_lerobot_output_qwen3vlablation_nemotron_thinking_32k_with_reasoning_effort
Dataset: ablation_nemotron_thinking_32k_with_reasoning_effort
This dataset was uploaded from /mnt/yulan_pretrain/mount/data_final_train/ablation_nemotron_thinking_32k_with_reasoning_effort/stage_1/tmp/.
CodeX-7M-Non-Thinking
Modotte
Note: This dataset is part of the lineup CodeX by Modotte. You can get lots of datasets in this same lineup, with the main focus on providing very high-quality datasets for model training and fine-tuning.
This dataset is curated from high-quality public sources and enhanced with synthetic data from both closed and open-source models. It serves as a strong foundation for instruction-based model tuning and fine-tuning, offering one of the most refined and extensive… See the full description on the dataset page: https://huggingface.co/datasets/Modotte/CodeX-7M-Non-Thinking.thinking-rollouts
thinking-rollouts
Unconstrained rollouts from thinking (chain-of-thought) models on DS-1000 and LiveCodeBench, CoT
saved verbatim alongside the final answer. Format per genlm/rollouts issue #5; schema is a superset
of temperature-sweep-data.
Hive-partitioned Parquet, thinking_mode folded into the model tag:
rollouts/domain=<dataset>/model=<tag>/temp=<temp>/data.parquet (tags like qwen3-8b-think,
qwen3-1.7b-nothink). 100 samples/instance.
Columns: model, thinking_mode, temp… See the full description on the dataset page: https://huggingface.co/datasets/samuki-hf/thinking-rollouts.openthoughts4-code-9168-prompts-qwen3-30b-a3b-thinking-2507-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-30B-A3B-Thinking-2507 (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-30b-a3b-thinking-2507-n16-flattened-logprobs-k16.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.llm-jp-4-33b-thinking-dpo-data
llm-jp-4-33b-thinking-dpo-data
Overview
This dataset is a Direct Preference Optimization (DPO) dataset used to train llm-jp-4-33b-thinking.
It is constructed by pairing multiple candidate responses for a given prompt and selecting preferred (chosen) and non-preferred (rejected) responses. The splits reasoning_low, reasoning_medium, and reasoning_high correspond to different reasoning effort settings used during response generation.
The fields chosen_analysis… See the full description on the dataset page: https://huggingface.co/datasets/llm-jp/llm-jp-4-33b-thinking-dpo-data.llm-jp-4-thinking-sft-data-chatmlllm-jpのデータセットllm-jp-4-thinking-sft-dataを、
ChatML形式に変換したものです。
ライセンス
各サンプルのライセンスは、元データセットカードに記載された各データソースのライセンスに従います。
本リポジトリは、元となったデータ全体に対して新たなライセンスを付与するものではありません。
利用する場合は、対応する元データソースのライセンス条件を確認してください。
llm-jp-4-8b-thinking-dpo-data
llm-jp-4-8b-thinking-dpo-data
Overview
This dataset is a Direct Preference Optimization (DPO) dataset used to train llm-jp-4-8b-thinking.
It is constructed by pairing multiple candidate responses for a given prompt and selecting preferred (chosen) and non-preferred (rejected) responses. The splits reasoning_low, reasoning_medium, and reasoning_high correspond to different reasoning effort settings used during response generation.
The fields chosen_analysis, chosen_final… See the full description on the dataset page: https://huggingface.co/datasets/llm-jp/llm-jp-4-8b-thinking-dpo-data.llm-jp-4-32b-a3b-thinking-dpo-data
llm-jp-4-32b-a3b-thinking-dpo-data
Overview
This dataset is a Direct Preference Optimization (DPO) dataset used to train llm-jp-4-32b-a3b-thinking.
It is constructed by pairing multiple candidate responses for a given prompt and selecting preferred (chosen) and non-preferred (rejected) responses. The splits reasoning_low, reasoning_medium, and reasoning_high correspond to different reasoning effort settings used during response generation.
The fields chosen_analysis… See the full description on the dataset page: https://huggingface.co/datasets/llm-jp/llm-jp-4-32b-a3b-thinking-dpo-data.hermes-function-calling-thinking-V1highlevel_thinking_with_grounding_annotation_split1000_v3