Dane dla AI

Data for AI · id data

Pozyskiwanie, przygotowanie, selekcja i kontrola danych używanych przez modele.

Sourcing, preparing, selecting and controlling the data models are built on.

In one sentence

To przygotowanie informacji i przykładów, z których modele uczą się i na których są sprawdzane.

This prepares the information and examples that models learn from and are tested on.

Problem it solves

Błędne, przypadkowe lub niepełne przykłady mogą nauczyć model złych nawyków.

Incorrect, irrelevant or incomplete examples can teach a model the wrong patterns.

What it does

Zbiera materiały, sprawdza ich pochodzenie, poprawia błędy i wybiera przykłady odpowiednie do nauki oraz osobnego sprawdzianu.

It collects material, checks its origins, fixes problems and selects examples for learning and a separate assessment.

Everyday example

Przygotowujesz zdjęcia produktów do rozpoznawania uszkodzeń: oznaczasz wady, usuwasz powtórzenia i odkładasz część zdjęć do sprawdzenia modelu.

You prepare product photos for damage detection, label defects, remove duplicates and set some photos aside to test the model.

This concerns you if you…

• chcesz nauczyć model pracy na własnych materiałach
• zastanawiasz się, skąd biorą się błędy modelu
• want a model to learn from your own material
• wonder where a model's mistakes come from

Examples

• Crawling · filtrowanie jakości · mieszanki danych · etykietowanie · dane syntetyczne
• Crawling · quality filtering · data mixtures · labelling · synthetic data

Subtopics

• Atrybucja i wycena danych treningowych (Training data attribution and valuation): Methods estimating the influence of training examples on predictions or their value for model quality, including influence functions, TRAK and Shapley values.
  • Funkcje wpływu i śledzenie przykładów treningowych (Influence functions and training example tracing): Gradient-based approximations of a training example's influence on a test prediction: classical influence functions, TRAK and scalable variants.
  • Wycena danych wartością Shapleya (Data Shapley) (Data Shapley valuation): The cooperative-game Shapley value applied to training examples, estimated by Monte Carlo sampling or gradient methods.
• Czyszczenie i jakość (Data cleaning and quality): Removing errors, duplicates and unwanted examples.
  • Deduplikacja i dekontaminacja (Deduplication and decontamination): Removing duplicates and near-duplicates, and cutting training examples that overlap with test sets.
    • Deduplikacja semantyczna (po embeddingach) (Semantic deduplication with embeddings): Removing semantic near-duplicates by clustering embeddings from a pretrained model and pruning pairs above a similarity threshold.
  • Ekstrakcja tekstu ze stron i PDF (Text extraction from HTML and PDFs): Pulling the actual content out of raw pages and documents: removing menus, footers and ads, turning PDFs (including scans) into clean text while keeping structure.
  • Filtrowanie jakości (heurystyki i klasyfikatory) (Quality filtering with heuristics and classifiers): Dropping weak documents before training — first with simple rules, then with a model that scores how valuable a text is (e.g. educationally).
  • Filtrowanie szkodliwych treści tekstowych w danych treningowych (Harmful text filtering in training data): Classifier- and rule-based removal of toxic, hateful and sexual content from text pretraining corpora.
  • Niezbalansowane klasy i wartości odstające (Class imbalance and outliers): Spotting and handling data where one class dominates or unusual examples appear, before they distort training and evaluation.
  • Wykrywanie błędnych etykiet (Label error detection): Automatically finding examples that probably carry a wrong label, using model predictions, instead of reviewing the whole dataset by hand.
• Dane syntetyczne i augmentacja (Synthetic data and augmentation): Creating or transforming examples.
  • Augmentacja danych (obraz i tekst) (Data augmentation for images and text): Creating new examples by transforming existing ones — rotating and cropping images, paraphrasing and back-translating text.
  • Generowanie danych instrukcyjnych przez LLM (LLM-generated instruction data): A model generates instructions, inputs and answers from a few seed examples; after filtering weak and near-duplicate ones they are used for fine-tuning.
  • Syntetyczne dane do pretrainingu (przepisywanie, podręczniki) (Synthetic pretraining data: rephrasing and textbooks): Rewriting web text into a more useful form or generating textbook-style material as a way to improve pretraining efficiency. Outcomes depend on generation quality, filtering, and how synthetic data is combined with real data.
  • Syntetyczne dane tabelaryczne (Synthetic tabular data generation): Generating rows for mixed-type tables with generative models; CTGAN conditions generation to better represent multimodal continuous columns and rare categories.
  • Syntetyczne trajektorie i dane wywołań narzędzi dla agentów (Synthetic agent trajectories and tool-use data): Automatic generation and verification of queries, function calls and multi-step agent trajectories in environments and APIs for model fine-tuning.
  • Weryfikacja i filtrowanie danych syntetycznych (Verifying and filtering synthetic data): Quality control for model-generated data through rejection sampling and automatic verification of answers, code execution or tool calls.
  • Wzbogacanie próbek (Sample enhancement): Improving generated examples: checking answers, picking the best of several, making tasks harder.
  • Załamanie modelu na danych syntetycznych (Model collapse from synthetic data): The risk that models trained recursively on content generated by earlier models lose information about the original data distribution, beginning with rare cases.
• Etykietowanie i feedback (Annotation and feedback): Producing reference answers and judgements.
  • Aktywne uczenie (active learning) (Active learning): The model picks which examples are worth labelling next, to reach good quality with fewer labels.
  • Etykietowanie z pomocą LLM (LLM-assisted annotation): Using language models to produce labels, rationales and pairwise judgements, with quality checked on a human-reviewed sample.
  • Słaba superwizja i funkcje etykietujące (Weak supervision and labeling functions): Instead of labelling by hand, you write many simple, imperfect rules and a model combines their votes into a single better label.
  • Zbieranie danych preferencji (porównania odpowiedzi) (Preference data collection (response comparisons)): Designing and running collection of pairwise comparisons or response rankings from people as data for reward models and preference optimisation.
  • Zgodność anotatorów (Inter-annotator agreement): Measuring how consistently independent annotators assign labels while accounting for chance agreement. A low score requires diagnosis: it may reflect unclear categories, task subjectivity, class distribution, or an unsuitable agreement coefficient.
• Kuracja danych multimodalnych (Multimodal data curation): Filtering, segmenting and (re)captioning image–text, video and speech data before pretraining multimodal and generative models.
  • Filtrowanie par obraz–tekst (Image–text pair filtering): Selecting image–text pairs from a web pool by embedding similarity, rules and classifiers, judged by the quality of the trained model.
  • Kuracja danych wideo (Video data curation for generative models): A multi-stage pipeline: cut detection, clip captioning and filtering by motion, on-screen text, aesthetics and image–text alignment.
  • Przygotowanie danych mowy z nagrań „z życia” (In-the-wild speech data curation): A pipeline of normalisation, source separation, diarization, VAD segmentation, automatic transcription and quality filtering for large speech corpora.
  • Syntetyczne podpisy do obrazów (recaptioning) (Synthetic image recaptioning): Generating new, dense image descriptions with a vision-language model and training on them instead of or alongside original captions.
• Pipeline’y i wersjonowanie (Data pipelines and versioning): Repeatable data preparation across experiments.
  • Przetwarzanie danych na dużą skalę (Large-scale data processing): Building repeatable pipelines that read, filter, deduplicate and tokenise trillions of tokens in parallel across many machines.
  • Strumieniowe, shardowane zbiory treningowe (Streaming sharded training datasets): Organising samples into sequentially read shards and streaming them from disk or object storage into training.
  • Walidacja i testy danych (Data validation and testing): Automatically checking that incoming data has the expected schema, ranges and distributions before it reaches training or a production model.
  • Wersjonowanie zbiorów danych (Dataset versioning): Keeping versions of data the way you keep versions of code, so every model can be tied to the exact dataset it was built on.
• Podziały i wyciek danych (Data splits and leakage): Keeping study material apart from an honest exam.
  • Rodzaje wycieku danych i ich wykrywanie (Types of data leakage and their detection): Recognising common mistakes that let a model peek at the answers: preprocessing train and test together, duplicates, features unavailable in practice, leakage from the future.
  • Strategie podziału: losowy, czasowy, grupowy (Split strategies: random, temporal, grouped): Choosing how to split data for the situation: by time for forecasting, by user or patient when examples are related to each other.
• Pozyskiwanie i pochodzenie (Data sourcing and provenance): Where the data comes from and whether we may use it.
  • Licencje i zgody na użycie danych (Data licensing and consent): Checking the terms under which data may be collected and used for model training: dataset licences, terms of service, consent, and opt-out signals. Robots.txt rules govern crawler access but are not themselves a licence or legal authorization to use the data.
  • Pozyskiwanie danych z sieci (crawling) (Web crawling for training data): Collecting web pages at scale as raw training material: using existing web snapshots or running your own crawler that respects site rules.
  • Wyszukiwanie i ocena gotowych zbiorów danych (Finding and assessing public datasets): Using dataset hubs, catalogues and their metadata (dataset cards, Croissant) to select, inspect and load data.
  • Śledzenie pochodzenia danych (lineage) (Data lineage and traceability): Recording where each part of a dataset came from, who created it and how it was processed, so it can be audited, fixed or removed.
• Selekcja i proporcje danych (Data selection and mixing): Deciding which examples, and in what proportions, enter training to improve model quality or budget efficiency when data sources differ in quality and usefulness.
  • Dobór mieszanki danych (DoReMi, RegMix) (Data mixture optimisation): Setting how much each source (web, code, books, science) contributes to training, e.g. using small proxy models.
  • Kolejność danych i uczenie z programem nauczania (curriculum learning) (Data ordering and curriculum learning): Controlling the order and pace of data: curricula from easier to harder examples and automatic self-paced learning.
  • Selekcja danych do dostrajania (Selecting data for fine-tuning): Choosing a small, carefully curated set of SFT examples instead of a large random one — by hand or with a method that measures each example's influence on the target task.

Sources

• Full Stack Deep Learning (2022, archiwum) · reference: Pełny cykl systemu ML: dane, testy, wdrażanie, monitoring, zespoły i etyka.
• Stanford CS336 · Language Modeling from Scratch (2025) · reference: Budowa modeli, dane, trening, systemy, inferencja i post-training.
• Introduction to Data-Centric AI (MIT IAP course, 2023–2024) · MIT CSAIL (prowadzący na stronie tylko z imion: Anish, Curtis, Jonas)
• A Survey of LLM × DATA · Xuanhe Zhou, Junxuan He, Wei Zhou, Haodong Chen, Zirui Tang, Haoyu Zhao, Xin Tong, Guoliang Li i in. (17 autorów)