Alternatives to Velvet

Applied research lab that captures expert reasoning and structures it into SFT, RL rubric, agent, and computer-use training data for frontier models.

Cortex AI supplies real-workplace egocentric video and robot trajectory data for training embodied AI models.

OpenCompass (司南) benchmarks LLMs, VLMs, and AI4S models against 100+ open evaluation datasets with published, dated leaderboards

Offline AI image annotation for YOLO and VOC datasets — local SAM2, SAM3 and YOLO-World, one-time purchase from $49.

Snorkel AI builds expert training data, evals, and runnable environments for frontier models and agents.

Silico is Goodfire's interpretability agent for understanding, debugging, and controlling the internals of your AI models

Weights & Biases tracks ML experiments and traces LLM apps so teams can ship AI models faster

Open-source, multimodal data annotation platform for image, video, 3D point cloud, and audio labeling—self-hosted, cloud, or fully managed

Human-recorded computer-use datasets that teach AI agents to operate real software the way people do.

MathArena is a free LLM math benchmark leaderboard scoring frontier models on ArXivLean proofs, BrokenArXiv, and ArXivMath competition

GameEngineBench benchmarks and UE5 training data for AI agents that write real game-engine code.

Open-source Python framework for generating grounded synthetic datasets from real sandboxed tool execution traces.

Potato is a free, open-source annotation tool for text, image, audio, video, and AI agent trace labeling — all configured in YAML.

BioStack Platforms supplies structured clinical datasets and RL environments for healthcare AI teams.

Web3 data network paying contributors in crypto for text, audio, video, code, and mapping data that trains AI models.

Fiddler AI is an enterprise AI control plane for agent observability, guardrails, and governance across the agentic lifecycle.

Raiinmaker builds custom, rights-cleared video datasets and runs real-time human evaluation for generative AI video models.

Simulation-first evaluation and training infrastructure for AI agents, built on Digital World Models.

Early-stage San Francisco robotics lab building policies for robots that do real work, and hiring its first researchers.

A human data platform connecting AI teams and researchers to 300,000+ verified participants for preference tuning, safety evals, and online studies.

A free 70,000-image Zalando clothing dataset built as a drop-in replacement for MNIST.

Daft is an open-source multimodal data engine for turning video, images, audio, and sensor data into training-ready datasets.

Audio datasets built like research artifacts for teams training speech recognition, translation, synthesis and voice-interaction models

Pre-trained finance AI models and API infrastructure for banks, insurers, and lenders, moving into agentic workflow orchestration.

Managed training data, RL environments, and safety red-teaming for teams building AI agents, coding models, and LLM/VLM pipelines.

Outlier AI pays subject-matter experts to train AI models remotely on flexible, project-based work with weekly pay.

Open-source Python toolkit that auto-labels images with big foundation models, then trains small fast vision models you own.

Web data platform for AI training and agentic web access — pre-collected datasets, live scraping APIs, and a proxy network under one account.

Custom audio datasets for training and evaluating voice AI, scoped to your model's spec.

A downloadable Python annotation tool and web app you run yourself to build training and evaluation data for custom AI, ML and NLP models.