dataset-curation

SDK for robotics teams to verify the quality of their data used for AI model training.

Content for workshops on computer vision @ HPI's AI Service Center

Turn internet-scale video into egocentric and exocentric training data, curated and annotated by agents on the Memories.ai datalake

HyperView curates datasets and provides model introspection in hyperbolic and Euclidean geometries.

Implementation of TSDS: Data Selection for Task-Specific Model Finetuning. An optimal-transport framework for selecting domain-specific and task-specific training data to improve LLM finetuning and instruction tuning.

Harnessing Large Language Models for Curated Code Reviews

Annotate any LeRobot dataset with a swarm of agents. Turn raw demos into training-ready data for robotic foundation models.

MedPMC: a high-fidelity medical image–text curation framework from biomedical literature for multimodal foundation models, with released pipelines, curated data, benchmarks, and pretrained models.

Image description/tagging tool

Semantic pipes for your terminal - PDFs, images, audio, video, and text through Unix verbs that understand

NAACL 2025 | How Can We Diagnose and Treat Bias in Large Language Models for Clinical Decision-Making?

[ACL 2024 (Findings)] ICC: Quantifying Image Caption Concreteness for Multimodal Dataset Curation

This is my personal home rig for serious LLM experimentation. I built it to test models head-to-head, create custom evaluation rubrics, automatically improve prompts based on the previous run’s results, and generate high-quality synthetic training data. Everything runs locally first (Ollama by default), with optional cloud support. logged locally.

An image deduplication GUI, made for image generation models dataset deduplication using CLIP.

AI-assisted curation and organization for large media datasets

Turn real traffic/traces into a versioned, deduped, coverage-tracked LLM regression suite — a curation accelerator with a human-confirmed golden gate.

Your dataset discovery and curation buddy.

Local image-dataset curation and captioning workspace for training-set builders — cull, group, tag, caption, export

Dynamic cluster-based data sampling for efficient and long-tail-aware vision-language model pre-training.

Multi-axis VLM-as-judge data curation for robot demonstrations. Five specialist critics (visual / kinematic / task / strategy / safety) score every LeRobot episode with rationale and timestamp evidence. Model-agnostic — any OpenAI-compatible vision endpoint.