The Dataset Finder blog
Working notes on the open datasets worth your time — what each one contains, how it is licensed, and the models it actually suits. No affiliate links, no padding.
Latest · Medical imaging
Expert contours differ by measurable amounts, so the reference standard is a design choice. LIDC’s four-reader protocol, STAPLE, report-mined labels and DICOM pitfalls.
Nothing in that topic here.
Image, video and vision-language collections — what each contains, how it is licensed, and the models it actually suits.
Expert contours differ by measurable amounts, so the reference standard is a design choice. LIDC’s four-reader protocol, STAPLE, report-mined labels and DICOM pitfalls.
Computer visionThe 1,000-per-class rule dates from ImageNet. Pretrained models now reach 77% to 85% ImageNet accuracy from 10 per class. What learning curves show, and how to forecast your own number.
Autonomous drivingHow KITTI, nuScenes, Waymo and Argoverse 2 define 3D boxes, headings, keyframes and track IDs, where they disagree, and what point-wise labels cost: 4.5 hours per residential tile in SemanticKITTI.
VideoWhat VATIC, SAM 2, MOT16 and TAO measured about keyframe spacing, interpolation, identity rules and frame rates, and how to check the tracks a labelling team delivers.
Computer visionHigh-resolution source images for upscaling — and why synthetic downsampling limits real-world performance.
SportsPlayer detection, action recognition and game state — a small category with unusually well-defined tasks.
RoboticsManipulation trajectories, point clouds, meshes and embodied environments — and the inherited data terms to check behind a permissive licence.
MultimodalImage-text pairs, why negatives come from your batch rather than your labels, and what that means for hardware.
Generative AICaption-image pairs, aesthetic subsets and the licence questions that matter more here than anywhere else.
Computer visionBounding-box data across general, traffic, aerial and specialist domains — with the class balance and annotation quality caveats that decide results.
Computer visionMonocular and RGB-D depth data — and the difference between sensor depth, stereo depth and synthetic ground truth.
MultimodalImage-text pairs and visual instruction data — what each set contains, how it is licensed, and which models it suits.
AgricultureLeaf disease, weed discrimination, crop species, yield forecasting and livestock weight — and why the clean ones mislead.
Document AIText recognition, layout analysis, table extraction and form understanding — the four distinct tasks people call OCR.
Computer visionCOCO, ImageNet, Open Images, Pascal VOC, CelebA, LFW, ScanNet and NYU Depth V2 — the ones your tooling already assumes.
Computer visionPixel-level class labels for scenes, medical images and aerial views — and why mask quality varies so much more than box quality.
IndustrialCracks in concrete and bridges, road damage, construction-site safety and waste sorting — and why, for defect inspection, anomaly detection often beats classification.
MultimodalImage-question-answer triples — and the language-prior problem that lets models score well without looking.
Computer visionThe directory layout, the annotation conventions that silently break, and how few images you actually need per class.
Document AIScene text, invoices, document parsing and structured forms — the cases where recognition is still hard.
Medical imagingChest x-rays, knee grades, 3D CT and brain MRI — with the licence conditions that decide whether you can ship what you build.
Autonomous drivingSegmentation, signs, lights, plates and tracking — including which ones are permissive enough to put in a product.
Computer visionSingle-label classification data — the simplest vision task and the best place to establish a baseline.
MultimodalImage-description pairs from short alt-text to paragraph-length detail — and why caption length determines what your model produces.
Generative AIWhat a few dozen images can and cannot teach, how caption style controls the trigger, and the rights question that is unavoidable here.
Computer visionPer-object masks rather than per-class, why crowd regions are encoded differently, and the cheaper routes to getting masks at all.
Computer visionThe task where you cannot enumerate the positive class — and where the method follows from that rather than from the dataset size.
Faces & peopleFace attributes, detection, captioning, hand gestures and activity — and the consent and regulatory questions that apply to all of it.
NatureSpecies identification, biodiversity surveys and digitised collections — with the long-tail problem that defines them.
Computer visionKeypoint data for human pose — body, hands and faces, in 2D and 3D — with the skeleton-definition problem that blocks combining them.
GeospatialLand cover, cloud masking, oriented detection, super-resolution and wildfire risk — most under licences that allow commercial use.
VideoAction recognition, temporal grounding, captioning and video generation — with the storage and licensing realities of working with video.
Computer visionPromptable segmentation changes what training data means — and often means you should be generating masks with the model rather than training it.
RetailProduct catalogues, reviews, visual search and recommendation data — and why the public sets rarely match a real catalogue.
Text corpora, speech and audio sets, code, and the structured data that still runs most production machine learning.
Museum and art collections, comics and cartoons, film, book and game metadata, and the miscellany that does not fit a category — much of it under public-domain or permissive terms.
Audio & speechFrom 1,000 hours of clean read speech to 10,000 hours of audiobooks, podcasts and YouTube and 29,300 validated hours in 295 languages, plus sound events, emotion and speaker identity.
CodeSource corpora and instruction data for programming — plus the licence question that is unavoidable with code.
AgentsScreenshots, element trees, action traces and web task benchmarks — the data behind models that operate interfaces.
NLPPolarity, fine-grained emotion, intent and topic labels — and why cross-domain transfer disappoints in this category.
NLPParallel corpora, multilingual web text and the low-resource language problem — what is available and how uneven it is.
SafetyToxicity, harmful content, jailbreak attempts and machine-generated-text detection — the sets used to train and test guardrails.
NLPEntity-labelled text across general, legal, Romanian and invoice domains — and the tag scheme problem.
Audio & speechPaired audio and transcripts — with the sampling rate, accent and recording-condition checks that predict real performance.
TabularForecasting data — retail demand, network traffic, parking, finance and intensive care — and the leakage mistake that makes results look excellent.
NLPDocument-summary pairs across news, legislation, Reddit posts and philosophy papers — and why reference summaries are a weak target.
Audio & speechNon-speech sound — environmental audio, acoustic scenes, gunshots and clinical sounds, with the timing granularity to check before you choose.
TabularRatings, review-level feedback and catalogue metadata — with the offline-evaluation problem that makes published results hard to trust.
Audio & speechActed versus natural speech, categories versus dimensions, and why in one well-known corpus listeners agreed with the intended emotion only four times in ten.
Audio & speechPristine consistency instead of realistic variety, the recording properties that matter, and the consent question specific to synthesis.
Audio & speechPaired audio and text, the sampling-rate requirement that is not optional, and why your transcription convention is a product decision.
TabularStructured rows and columns — the format where gradient boosting is still the baseline to beat for deep learning, and why that matters for your data budget.
NLPParallel corpora and their alignment quality — plus the low-resource reality behind the headline language counts.
Finance & legalEarnings calls, loan and complaint records, legislation and case law — with the compliance questions that come attached.
NLPTopic, intent, toxicity and sentiment labels — and how to tell whether a public set matches your taxonomy.
Audio & speechWhy many utterances per speaker across separate sessions is the property that matters, and why overlapping speech is where diarisation actually fails.
NLPEncoder models need a completely different dataset shape from generative LLMs — and for classification they are still often the better choice.
What your data has to look like for each framework and model family, and the format bugs that fail silently.
torchtune is no longer actively maintained. What it was good at, what its dataset format looked like, and where to move a project that depends on it.
Fine-tuningThe exact dataset shapes Unsloth accepts, the EOS token bug that makes models generate forever, and how to find data worth training on in the first place.
Fine-tuningZero-code fine-tuning with a web UI, and the most explicit column-mapping system in the ecosystem — how to register your own data correctly.
Fine-tuningThe chat template, the acceptable-use terms that travel with the weights, and what the base model is already good enough at that you should not be training it.
Fine-tuningSeveral Mistral releases are Apache 2.0, which makes this family unusually clean for commercial work — but not all of them are, and the templates differ.
Fine-tuningYAML-configured training, arbitrary key remapping, and the pre-tokenised and template-free formats that give you per-span control over the loss.
Fine-tuningWhat is actually published versus what is folklore, why rank matters less than people think, and how to tell from your loss curve that you have too little.
Fine-tuningThe Phi models were trained largely on curated and synthetic textbook-quality material. That shapes what they are good at and what your fine-tuning data should look like.
Fine-tuningA decision procedure rather than a list: work backwards from the behaviour you want, the stage you are at, and what you are allowed to ship.
Fine-tuningStrong non-English coverage, a wide family including vision and coding variants, and the thinking-mode templates that need separate handling.
Fine-tuningUnsloth builds directly on TRL's trainers, and Axolotl and LLaMA-Factory use TRL for preference training such as DPO, so for those methods TRL's dataset types are the real spec.
Fine-tuningApache 2.0 for Gemma 4 but Google's own terms for earlier generations, a distinct chat format, and very small models that make unusual projects practical.
Pretraining through preference tuning: what each stage consumes, at what scale, and how anyone measures whether it worked.
The famous 80% agreement came from GPT-4 on MT-Bench with ties removed. What it measured, the biases documented since, and how many human labels it takes to validate a judge.
LLM trainingOLMo 2 defines it as 5–10% of training FLOPs with the learning rate annealed to zero. What goes in the mix, and what it is worth.
BenchmarksMMLU, GSM8K, MATH, IFEval, HumanEval and the reasoning benchmarks — with the published scores that give you a sense of scale, and the contamination problem.
LLM trainingMost open post-training data is now machine-generated. The named teachers, the real scales, and the licence question that decides whether you can use any of it.
LLM trainingThe chosen/rejected contract, the real sizes of UltraFeedback and Tülu 3's preference mixes, and the hyperparameters that differ sharply from SFT.
LLM trainingSFT mixtures and reasoning data for fine-tuning language models — sizes, licences and what each one teaches.
LLM trainingFineWeb, Dolma, The Pile, DCLM, RedPajama and Common Pile — what each actually contains, how it was filtered, and which ones have a licence you can rely on.
BenchmarksThe evaluation sets models are measured on — what each probes, and why contamination makes public scores drift upward.
LLM trainingThe open SFT mixtures with their real row counts, and the three-orders-of-magnitude disagreement between the LIMA camp and the Tülu camp.
LLM trainingWhen instruction tuning cannot help because the model lacks the knowledge — how much domain text you need, how to avoid forgetting, and the embedding learning rate people miss.
LLM trainingInstructGPT's actual data scale — 13k, 33k, 31k prompts and 40 contractors — and why most teams now use DPO instead of a reward model and PPO.
BenchmarksThe highest-value artefact in your pipeline, and the one most teams skip. How to build a few hundred examples that make every later decision answerable.
LLM trainingReinforcement learning where a program checks correctness instead of a reward model. Tülu 3's exact prompt mix, and the tasks this does and does not work for.
Manipulation, teleoperation, world models and egocentric video — including which of it you are actually licensed to train on.
DROID, AgiBot World, BridgeData V2 and the pooled archives — scale, hardware, and which ones survive a commercial licence check.
Physical AIEgocentric human video, navigation, and the simulation assets — the data behind agents that move through environments rather than just grasp in them.
Physical AIThe open corpora for robots and embodied agents, their real scale, and the licence problem — one of the largest manipulation datasets in the world is non-commercial.
Physical AIEgo4D, Ego-Exo4D and EgoDex compared on what is paired with the video — and why the largest and most useful ones are the hardest to use commercially.
Physical AISixty datasets from 34 labs in one format, with a coarsely aligned action space. What it contains, what it trains, and why 'OXE is CC-BY' is not quite true.
Physical AIParquet for signals, MP4 per camera, JSON for schema — and why v3 stopped using filenames to find episode boundaries.
Physical AIOf the major world-model programmes, only a handful — Cosmos, V-JEPA 2 and the original Genie, plus AgiBot's Genie Envisioner — publish usable training-data figures. That asymmetry tells you something about the field.
Physical AIUnlimited volume and perfect labels against a transfer gap. How teams actually combine the two, and the licence trap in simulator scene assets.
Physical AIWhat it costs per task, the hardware people actually use, and the published finding that should reshape your collection plan.
Physical AIWhy navigation data looks nothing like manipulation data, where the scene assets come from, and the licence question that sits under the simulator.
How much data you need, whether to generate it, and how to tell good data from a lot of data.
Networks tolerate random label flips far better than structured human errors, and errors in the test set cost more than errors in training. What CIFAR-N and Northcutt et al. measured.
AnnotationCareful re-evaluations found active learning 2 to 4 times more label-efficient in one setting and no better than random in others. What decides which, and how to test it before you rely on it.
AnnotationChatGPT beat crowd workers at labelling tweets; on expert finance, law and biomedical tasks the best models scored 47.5% to 87%. Where LLM labels hold, where they fail, and how to check.
Data strategyWhat the studies say about label quality by workforce, where domain expertise has to stay close, the costs teams forget, and a decision matrix you can score.
AnnotationOne label, three or five: what Snow et al. and Snapshot Serengeti measured, when majority vote is enough, when Dawid–Skene earns its keep, and what each extra label buys.
AnnotationSAM cut time per mask from 34 to 14 seconds and SAM 2 cut video time per frame 8.4-fold. What those figures measured, where pre-labels fail, and how they bias the people correcting them.
AnnotationMeasured seconds per label for tags, points, boxes, polygons and masks, from COCO, Cityscapes and the click-supervision papers, and what each extra second buys a model.
AnnotationHow many checked items it takes to estimate an error rate, how to set an acceptance sample for a batch, and what gold items, consensus and expert review each catch.
AnnotationCompute Cohen’s kappa, weighted kappa and a 95% confidence interval for two annotators, with per-label agreement and what to fix when it is low.
AnnotationBuild labelling guidelines annotators can follow: label definitions, edge cases, format and quality bar. Export to Word or PDF and order the labelling.
Data strategyWhat to write down about every dataset you use, why it takes minutes at the time and is impossible afterwards, and the questions it answers.
Data strategyFrom phi-1's textbooks to Nemotron-CC's 1.9 trillion rephrased tokens — the published results, and the failure modes that do not show up in your loss curve.
Data strategyThe compute-optimal rule, the exact figures behind it, and why Llama 3 trained a 405B model on 15.6 trillion tokens instead of the 8 trillion the rule implies.
Data strategyA practical procedure: measure composition, compare against deployment, test subgroup performance, and know which gaps you can fix with data.
Data strategyFive published results at different scales, including the one that shows aggressive filtering can backfire.
Data strategyThe fields people actually need when deciding whether to use your data — and the ones most cards omit.
Data strategyWhat is published versus what is folklore, the three-orders-of-magnitude disagreement in the literature, and an experiment that answers it for your task in an afternoon.
Data strategyExact, near and cross-split duplicates — what each costs you, and the MinHash settings the major corpora actually published.
Labelling companies, platforms and dataset sources side by side, with links to the detailed comparison pages.
Nine self-hostable labelling tools compared on licence, data types, model-assisted labelling and release activity, checked against each repository on 31 October 2026.
VendorsThe four pricing models, the few prices vendors publish (checked 31 October 2026), and a cost model built from measured seconds per label to test any quote.
VendorsSixteen alternatives to Scale AI for RLHF, managed annotation and labelling platforms, with pricing models, who does the work, and who each one suits.
VendorsWhere to find training data beyond the Hugging Face Hub: dataset search engines, research repositories, open-data portals, domain hubs and licensed data.
Vendors22 data labelling (data labeling) companies and platforms in 2027, grouped by use case, with who does the work, pricing models and what to ask before you buy.
The clauses that decide whether you can ship a model, and the questions to take to a lawyer before you train.
Article 10 paragraph by paragraph, the Digital Omnibus changes now in force, the new dates for high-risk systems, and the records to keep for each dataset.
ComplianceWhat de-identified actually means, why imaging is harder than records, and the approval routes for data you cannot simply download.
LicensingCC0 through CC BY-NC-SA, MIT, Apache, ODC-By and the research-only terms — what each permits, and the two clauses that cause nearly all the trouble.
ComplianceLawful basis, purpose limitation, the right to erasure and what any of it means for a model's weights. The issues, not the answers.
LicensingCC0, MIT, Apache 2.0 and plain CC-BY — permissive, commercial-friendly, and without the copyleft obligation that catches people.
LicensingThe datasets with the fewest strings attached — no attribution, no share-alike, no commercial restriction. What CC0 does and does not settle.
LicensingClick-through agreements, credentialed repositories, signed data use agreements and institutional review — what each involves and how long it takes.
Annotation and storage formats field by field, with converters that do not silently corrupt your labels.
Which instruction-tuning format to use, how each one maps onto a chat template, and the loss-masking decision that quietly determines whether your fine-tune works.
FormatsOne text file per image: a class index and four normalised numbers per line. Simple enough to get wrong — here is the exact spec plus a tested COCO converter.
FormatsA conversion checklist that catches the silent failures: coordinate conventions, class index drift, lost negatives and dropped metadata.
FormatsOne XML file per image, corner-based boxes, a difficult flag that changes your metrics. Why older datasets and annotation tools still emit it.
FormatsWhat every key in a COCO annotation file means, how segmentation and bbox differ, and the silent mistakes that corrupt a training run without making a loader throw.
FormatsColumnar storage, predicate pushdown and row groups — what they buy you on a dataset you are about to train on, and when Parquet is the wrong choice.
Tens of thousands of curated datasets, searchable in plain language — with the licence and size checked before you download anything.