#datasets

Train/test contamination inflates benchmark scores quietly. Here's what it is, the three levels of overlap worth measuring, and how to check two files without writing a script.

A pre-training checklist for fine-tuning datasets: schema, roles, duplicates, leakage, length distribution, label balance and the rendered-template sp

The dataset formats you will meet when fine-tuning an open LLM, how chat templates turn them into tokens, and the formatting mistakes that quietly rui

Real labelled images for a narrow vision task are often scarce. Rendered or generated synthetic data can fill the gap — with real limits worth knowing

Exact duplicates are easy to catch. Near-duplicates — the same example paraphrased — quietly bias a fine-tune and are far more common than people expe

Generating training examples with a bigger model can fill a data gap fast, or quietly bake in that model's biases and errors. How to tell which you're

Two months ago a repository turned up in my corner of the internet: 3,358 stars, 744 forks, an MIT...

Researchers report that raising the fidelity of handheld human demonstrations removes the need for any robot teleoperation on the target task, with policies trained on handheld data alone reaching parity with robot-taught baselines on four two-armed

Hugging Face released The Stack v3, a code dataset of roughly 4.9 trillion tokens across 173 million repositories with source text embedded inline, ending the v2 arrangement where researchers had to fetch the actual file contents themselves.

If you work in cybersecurity research, threat hunting, or machine learning, you've almost certainly...

An honest walkthrough of an 18.6M-snapshot prediction-market price dataset: what you can backtest with it, what you can't, and why I'm telling you the limits up front.

Anything you build around perfume data — a dupe/clone finder, a fragrance shop's product enrichment,...