datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
PP
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.opd-kd-thinky-deepmath-completions
train_rl Completion Logs
This dataset contains the on-policy generations produced during RL training
with train_rl.
Training details
Key
Value
Algorithm
OPD
Model (student)
HuggingFaceH4/KD-Thinky
Model (teacher)
Qwen/Qwen3-8B
Prompt dataset
HuggingFaceH4/DeepMath-103K
Group size
4
Max completion tokens
4096
Temperature
1.0
Learning rate
0.0001
model_revision
v00.08-step-000003125
dataset_configtrl_all
lora_rank
128
opd_kl_coef
1.0… See the full description on the dataset page: https://huggingface.co/datasets/kashif/opd-kd-thinky-deepmath-completions.agent-think-tool_use
Agent Think Tool Use
Датасет многошаговых агентных сессий для дообучения моделей работе с кодом, инструментами и инженерными задачами. Записи содержат пользовательские требования, комментарии агента во время работы, decision summaries, вызовы инструментов, результаты запусков, обработку ошибок и финальную проверку.
Каждый shard представляет отдельную связанную сессию, а не отдельный вопрос и ответ. Данные охватывают исследование задачи, работу с документацией, проектирование… See the full description on the dataset page: https://huggingface.co/datasets/ru-dataset/agent-think-tool_use.Dolci-Think-SFT-translated
Dolci-Think-SFT-translated
Machine translations of the Dolci-Think-SFT-32B dataset, produced with gemma-4-31B-it. The samples selected for translation are those where content_quality == "excellent" according to the propella annotations.
Columns
Each row is a translated conversation plus the result of a post-translation quality filter:
id — source record id.
messages — the translated conversation (list of {content, role}).
filter_pass — true if the row passed… See the full description on the dataset page: https://huggingface.co/datasets/openeurollm/Dolci-Think-SFT-translated.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.Dolci-Think-RL-7B
Dolci-Think-RL-7B
Dataset Summary
Dolci-Think-RL-7B is the reinforcement learning dataset used to train the Olmo-3-7B-Think model.It contains 102,014 prompts designed to elicit deep reasoning across:
Math
Coding
Precise Instruction Following
General Chat
It blends high-quality curated sources with filtering designed for deliberate reasoning.
Dataset Composition
Total Samples: 102,014
Original Dataset Contribution… See the full description on the dataset page: https://huggingface.co/datasets/allenai/Dolci-Think-RL-7B.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.browsecomp-qwen35-35b-a3b-think
browsecomp-qwen35-35b-a3b-think
Deep research agent evaluation on data/browsecomp.jsonl (normal split).
Results
Metric
Value
pass@4
43.0%
avg@4
24.8%
Trajectory accuracy
24.8% (1258/5064)
Questions
1266
Trajectories
5064 (4 per question)
Avg tool calls
41.1
Full conversations
❌
Model & Setup
Model
Qwen3.5-35B-A3B
Judge
gpt-4o
Max tool calls
50
Temperature
0.7
Blocked domains
huggingface.co
Tool… See the full description on the dataset page: https://huggingface.co/datasets/rl-rag/browsecomp-qwen35-35b-a3b-think.Soofi-Think-SFT-10B-multilingual
ReasonXL: A Multilingual Cross-Domain Reasoning Corpus
ReasonXL is a large-scale multilingual reasoning corpus spanning 5 languages and ~44B tokens in total. It is designed to support supervised fine-tuning of reasoning models with in-language chain-of-thought traces across diverse technical domains.
Data Generation
English source samples were drawn from 10 existing reasoning datasets, filtered and quality-annotated using ellamind/propella-1-4b, and then translated into… See the full description on the dataset page: https://huggingface.co/datasets/toroe/Soofi-Think-SFT-10B-multilingual.thinking-benchmark-90
Thinking Benchmark
A calibration pool of 90 competition-mathematics problems assembled to study how output / reasoning-trace length varies with problem difficulty across frontier language models. Part of the Cost of Overthinking research project.
Dataset at a glance
Source
n
Difficulty
Contamination risk
AIME 2026
29
3–5
low
OlymMATH
41
4–6
medium
HMMT February 2026
12
4–5
low
MATH-500
5
2–3
high
FrontierMath-style
3
6
medium
Difficulty is… See the full description on the dataset page: https://huggingface.co/datasets/tyrtleli/thinking-benchmark-90.highlevel_thinking_with_grounding_annotation_split1000_v3_merged_promptsopenthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-30B-A3B-Thinking-2507 (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 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-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16.4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.kv-tags-explained
4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.kv-tags-explained
A pre-tokenized, tag-wrapped variant of JackHsieh/4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.
The source thoughts are two parts: a <think> block, then a few dense sentences that the
generator wrapped in a literal <thought>…</thought> block (plain text, not special tokens).
Only the part after </think> becomes the VALUE, with every <thought> / </thought> tag
removed… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.kv-tags-explained.4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids
4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids
Thoughts for next-token prediction on k=8 token chunks of JackHsieh/statML-arxiv-40M-20M, generated by
Qwen/Qwen3-4B in thinking mode.
The prompted task: reason about what comes IMMEDIATELY next — the next k=8 tokens after the
cut — and answer with a few dense, declarative sentences (about two to four) inside a
<thought>…</thought> block, focused on the exact state at the cut and what the local
grammar… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/4B-think-reason-only.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.details_Qwen__Qwen3-30B-A3B-Thinking-2507_v2
Dataset Card for Evaluation run of Qwen/Qwen3-30B-A3B-Thinking-2507
Dataset automatically created during the evaluation run of model Qwen/Qwen3-30B-A3B-Thinking-2507.
The dataset is composed of 116 configuration, each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest… See the full description on the dataset page: https://huggingface.co/datasets/OALL/details_Qwen__Qwen3-30B-A3B-Thinking-2507_v2.4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.kv-tags-explained
4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.kv-tags-explained
A pre-tokenized, tag-wrapped variant of JackHsieh/4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.
The source thoughts are two parts: a <think> block, then a single paragraph of dense
reasoning about the immediate continuation. Only the part after </think> — the paragraph —
becomes the VALUE. The reasoning inside the think block is dropped.
The VALUE is capped at 512 tokens.… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.kv-tags-explained.4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids
4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids
Thoughts for next-token prediction on k=8 token chunks of JackHsieh/statML-arxiv-40M-20M, generated by
Qwen/Qwen3-4B in thinking mode.
The prompted task: reason about what comes IMMEDIATELY next — the next k=8 tokens after the
cut — and answer with a single unconstrained paragraph of dense reasoning, focused on the
exact state at the cut and what the local grammar, notation, or argument forces next. Both the… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/4B-think-paragraph.stride-train1-test32.k-8.statml-arxiv.qwen3-ids.dolci-think-translation-tests
Dolci-Think Translation Tests
This dataset contains the inputs, outputs, reconstruction records, automatic
diagnostics, and retained COMET-QE evidence from a set of controlled translation
experiments over 340 English messages selected from
allenai/Dolci-Think-SFT-7B.
It is intended for inspecting the experiments and reproducing analyses, not as a
ready-made translation training set.
The central complication is that one source document does not always correspond
to one model… See the full description on the dataset page: https://huggingface.co/datasets/RfKnowledge/dolci-think-translation-tests.kupe-thinkspark-270m-phase1-data
ThinkSpark-v2-350M — Phase-1 free-audio training data
Pre-encoded Mimi cb0 (12.5Hz semantic) tokens +
per-frame energy/f0 prosody, packaged for Phase 1 of ThinkSpark-v2-350M — teaching a
270M-parameter Gemma-3-based full-duplex floor-controller that a stream of Mimi audio
tokens carries language + prosody, before Phase 2 teaches it to referee turn-taking.
Sourced from free/open corpora (LibriSpeech, AI4Bharat Kathbath/Shrutilipi, IndicTTS,
Google FLEURS) via… See the full description on the dataset page: https://huggingface.co/datasets/anuj-inavlabs/kupe-thinkspark-270m-phase1-data.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-recall30-10m-historical-5t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-5t-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: 2.7000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 5-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-5t-think.dolci_think_rl_7b_messages_hybrid_450m_scoredharvey-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.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-5t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-5t-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: 3.7000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 5-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-5t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-5t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-5t-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: 1.9000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 5-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-5t-think.0.5M-thinking
0.5M Thinking Dataset
This dataset contains responses generated by MiniMax-M2.1 for user questions from the a-m-team/AM-DeepSeek-R1-Distilled-1.4M dataset (am_0.5M subset).
Dataset Description
The dataset captures both the extended thinking process and final answers from MiniMax-M2.1, with reasoning wrapped in <think> tags for easy separation.
Metric
Value
Examples
499,157
Total Tokens
3,732,749,397
Avg Tokens/Example
7,478
Source Dataset… See the full description on the dataset page: https://huggingface.co/datasets/PursuitOfDataScience/0.5M-thinking.Dolci-Think-RL-7B-Completions-SFT
Dolci-Think-Completions-SFT
Dataset Summary
Dolci-Think-Completions-SFT is a set of 5,031,398 completions(!!) from the Olmo-3-7B-Think-SFT model over the prompts considered when making Dolci-Think-RL.
These completions were mainly used to filter easy data, but we believe the completions may be useful in general.
It contains 636,095 high-quality prompts covering:
Math
Code
Precise Instruction Following
General Chat
Puzzles
Each split covers one of the above domains, and… See the full description on the dataset page: https://huggingface.co/datasets/allenai/Dolci-Think-RL-7B-Completions-SFT.open-thoughts-4-30k-code-qwen3-30b-a3B-thinking-2507-annotated-32768-tokens-n8
Open Thoughts 4 - Code (Qwen3-30B-A3B-Thinking-2507, 32K tokens, n=8)
This dataset contains code reasoning problems with 8 independent responses generated by Qwen3-30B-A3B-Thinking-2507.
Overview
Source: marin-community/open-thoughts-4-30k-code-qwen3-32b-annotated (prompts only)
Model: Qwen/Qwen3-30B-A3B-Thinking-2507
Temperature: 0.8
Max tokens: 32,768
Columns
Column
Description
instruction_seed
The code problem prompt
_source
Source dataset… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-code-qwen3-30b-a3B-thinking-2507-annotated-32768-tokens-n8.Soofi-Think-SFT-V2-secondhalf-DE