The AI category ships the file shapes machine-learning pipelines actually pass around — real formats with documented schemas, all fully synthetic. The NLP set is JSON Lines training data: labelled sentiment, token-level NER with BIO tags, chat fine-tuning in both the OpenAI and Anthropic message shapes (linked as a conversion twin), Alpaca-style instruction tuning, extractive QA in the SQuAD v2 structure with an unanswerable question, abstractive summarization, and an English–Spanish parallel corpus. The embeddings set carries the same 24 texts as an L2-normalised 16-dimensional vector set in three twinned encodings — JSON, Apache Parquet, and a raw NumPy .npy matrix — for testing vector stores and loaders. The vision set renders one detection scene and annotates it three ways — COCO JSON, YOLO text, and Pascal-VOC XML — so you can diff annotation converters against a known image. The evaluation set has benchmark results, a confusion matrix as CSV and JSON twins, an ROC curve, and a scikit-learn-style classification report; there's a templated prompt library, and a genuinely-valid tiny safetensors weight file (sample values, not a trained model). Every dataset uses fixed seeds and invented names, companies, and places — no real people or data.
The same synthetic conversations in the Anthropic Messages JSONL shape — a top-level system prompt plus a messages array of user and assistant turns. The format twin of the OpenAI file, for testing chat-format conversion.
A chat fine-tuning dataset in the OpenAI JSONL format — one conversation per line as a messages array with system, user, and assistant turns. Synthetic Q&A content. Paired with an Anthropic-format twin for testing format converters.
An extractive question-answering dataset in the SQuAD v2.0 JSON structure — titled articles with context paragraphs, questions, character-offset answers, and one deliberately unanswerable question. Synthetic content; a fixture for QA model training and SQuAD-format loaders.
An instruction-tuning dataset in the Alpaca JSONL format — 20 instruction / input / output triples covering small transformations, extraction, and factual answers. Fully synthetic; a fixture for supervised fine-tuning pipelines.
A token-classification dataset in JSON Lines — 16 tokenized sentences with aligned BIO tags for person, organisation, and location entities. All names, companies, and places are fictional. A fixture for NER model training and sequence-labelling tooling.
An English↔Spanish parallel corpus in JSON Lines — 20 aligned sentence pairs of everyday phrases. A fixture for training and evaluating machine-translation models and for testing UTF-8 handling of accented characters.
A labelled sentiment-classification dataset in JSON Lines — 24 short product-review-style sentences balanced across positive, negative, and neutral. Fully synthetic; a fixture for testing text-classification loaders, tokenizers, and JSONL parsers.
An abstractive-summarization dataset in JSON Lines — 15 short synthetic news-style documents each paired with a one-sentence summary. A fixture for training and evaluating summarization models and for testing JSONL ingestion.
A per-class classification report in the scikit-learn structure — precision, recall, F1, and support for each class plus accuracy and macro/weighted averages. A fixture for testing metric parsers and report renderers.
A 3-class confusion matrix as CSV — rows are the true class, columns the predicted class, cells the counts. Paired with a JSON twin for testing metric parsers and evaluation visualisers.
The same 3-class confusion matrix as JSON — a labels array plus a nested counts matrix. The structured twin of the CSV, for testing evaluation tooling.
A model-evaluation summary in JSON — per-task scores for a fictional model across sentiment, NER, summarization, translation, and QA, each with its metric and sample size. A fixture for testing eval dashboards and leaderboard importers.
An ROC curve as CSV — decision threshold with the corresponding false-positive and true-positive rates, monotonic from (0,0) to (1,1). A fixture for testing chart tools and AUC calculators.
Object-detection annotations for the scene in the COCO JSON format — images, categories, and per-object bounding boxes as [x, y, width, height]. Grouped with YOLO and Pascal-VOC twins for testing annotation-format conversion.
The same detection boxes in the Pascal VOC XML format — a per-image annotation with size, and one object element per box with pixel corner coordinates. The XML twin of the COCO and YOLO annotations.
The same detection boxes in the YOLO text format — one object per line as class id and box centre, width, and height normalised to 0–1. The format twin of the COCO and VOC annotations.
The class-name list for the detection scene, one label per line — index equals the zero-based line number, matching the YOLO class ids. A companion to the COCO/YOLO/VOC annotation files.
A simple rendered street scene with a person, a car, and a tree at known pixel coordinates — the image the COCO, YOLO, and Pascal-VOC annotation twins describe. A fixture for testing object-detection loaders and annotation converters.
A reusable prompt library in JSON Lines — 20 templated prompts for summarization, translation, extraction, code, and more, each with a task label, tags, and {curly-brace} placeholders. A fixture for prompt-management tools and JSONL parsers.
A set of 24 L2-normalised 16-dimensional text embeddings as JSON — each record pairs an id and its source text with a float vector. A fixture for testing vector stores, similarity search, and embedding loaders. Parquet and .npy twins included.
The same 16-dimensional embeddings as Apache Parquet — id and text columns plus one column per dimension. The columnar twin, for testing analytics engines and Parquet-based vector pipelines.
The embeddings as a raw NumPy array — a 24×16 float32 matrix in .npy format, loadable with numpy.load. The binary twin of the JSON and Parquet files, for testing tensor and matrix loaders.
A genuinely-valid safetensors file with two small float32 tensors (36 parameters total) — an 8×4 weight and a length-4 bias. The values are meaningless sample data, not a trained model; a fixture for testing safetensors loaders and weight inspectors.
We use Google Analytics. It sets no analytics cookies until you allow the Analytics category below. See our cookie policy.
