Codeset, a platform for training and evaluating agentic code models (codeset.ai)

Codeset publicly launched a platform and its first dataset, codeset-gym-python: a collection of 1,000 real-world Python bug-fixing tasks drawn from 291 permissively licensed repositories, each packaged with issue descriptions, commit history, and complete test suites. The platform exposes a Python SDK and REST-style API to spin up sandboxed, reproducible container environments per task, handle dependency installation, run commands, and execute verification jobs (test-suite runs) — all accessible via simple session and verification APIs. codeset-gym-python is available on Hugging Face and directly on Codeset from day one; the SDK is pip-installable and the service is pay-as-you-go with starter credits. For researchers and engineers building agentic code models, Codeset addresses major friction points: reproducibility, environment setup, and scalable verification. By turning complex build/test orchestration into API calls, it shortens experiment turnaround, makes large-scale training and evaluation pipelines more reliable, and enables closed-loop workflows where agents can iteratively propose fixes and be automatically validated. Planned expansions include additional languages (Java, JavaScript, Rust, Go, C++, C#), broader task types, integrations with other public datasets, custom dataset creation, and stronger, faster verifiers — all of which could significantly accelerate benchmarking, fine-tuning, and RL-style workflows for code agents.