OpenGVLab/GUI-Odyssey
GUI Odyssey is a comprehensive dataset for training and evaluating cross-app navigation agents on mobile devices, featuring 7,735 episodes across multiple devices, apps, and navigation tasks.
Awesome AI Agents › Training Datasets
always-further/deepfabric
DeepFabric is a Python pipeline for producing synthetic training data and evaluations for language models used in agentic systems. It generates samples that pair reasoning traces with tool-calling patterns, so a trained model learns to plan, call tools correctly and conform to strict schemas. The distinguishing mechanism is topic graph generation: instead of sampling prompts freely, the tool first builds a topic tree or graph of configurable depth and degree from a seed prompt, then generates samples per node. This is intended to keep coverage broad while staying anchored to the target domain and to avoid the redundancy that causes overfitting in other dataset generators. Output correctness is enforced on several levels. Constrained decoding and response validation keep samples inside the declared schema and variable constraints, and tool calls are actually executed inside isolated WebAssembly environments so execution correctness can be checked rather than assumed. Tool definitions can be imported directly from MCP server schemas and mocked automatically, used against real interfaces, or drawn from a standard set of common file operations such as list_files and read_file. The mocked execution path relies on a Spin service that the maintainers ship as a Docker image. Finished datasets can be uploaded automatically to Hugging Face and imported into training frameworks including TRL, Unsloth and Axolotl. After training, a built-in evaluation engine measures the model on unseen tasks derived from held-out splits, covering evaluation-only questions, answers and tool traces. DeepFabric can be used as a library, as a CLI, or through YAML configuration files with sections for shared LLM defaults, topic generation and sample generation, allowing a cheaper model for topics and a stronger one for samples. Conversation types include basic and chain-of-thought with an agent reasoning style, and agent mode is implicit once tools are configured. The project is Apache 2.0 and published on PyPI.
https://github.com/always-further/deepfabric
Also in Training Datasets
Dongping-Chen/GUI-World
GUI-World is a dataset and benchmark for evaluating and advancing multimodal large language models in understanding and interacting with dynamic graphical user interfaces, complemented by GUI-Vid, a fine-tuned VideoLLM for GUI video analysis.
neulab/MultiUI
MultiUI is a large-scale dataset and codebase for training and evaluating models on text-rich visual understanding of webpage user interfaces, featuring 7.3 million samples, pre-trained models, and comprehensive benchmark evaluations.