SmolDataEnvs has launched as an open collection of reinforcement-learning tasks designed to improve smaller models on code and data-science work. The project packages real tabular datasets, questions and verifiable answers so a training system can score whether an agent completed each task correctly.
The Hugging Face dataset card describes more than 5,500 tasks across the project. Its Harbor-formatted training split contains 5,000 hands-on data-analysis tasks, while separate evaluation and held-out test sets contain 144 and 250 tasks respectively. A supervised-fine-tuning split adds 4,677 verified example trajectories.
Built for repeatable training and evaluation
Each task drops an agent into a sandbox with a dataset and a question, then checks its result against a known answer. That structure is intended to make reinforcement learning practical without relying on subjective human ratings for every attempt.
The creators also provide standard dataset, Harbor and supervised-fine-tuning formats, allowing researchers to use the material without adopting a single training framework. The collection is released under the MIT license.
The launch announcement includes experiments with a 2-billion-parameter model and compares shuffled training with a curriculum ordered from easier to harder tasks. Those results come from the project team; broader usefulness will depend on independent runs, model choice and whether the benchmarks reflect real work outside the supplied environments.
