MLEng

Description

MLEng is an environment for evaluating machine learning engineering capabilities. Based on OpenAI's MLE-bench, agents are given Kaggle-style machine learning competitions and must develop, train, and submit ML models. Tasks cover diverse domains including classification, regression, and computer vision across real Kaggle competition datasets.

Capabilities

• Machine learning model development and training
• Data exploration and feature engineering
• Working with Kaggle competition formats
• Iterative model improvement and submission validation

Compute Requirements

Agents are given a sandbox with 4 CPUs and 8 GB RAM, with access to scientific Python libraries and ML frameworks.

License

MIT

Tasks

Two splits: mleng15 (14 tasks) and mleng100 (99 tasks). Total: 113 Kaggle-style ML competitions.

Reward Structure

MLEng is a multi-turn environment. Agent develops an ML model in the sandbox, produces a submission CSV at /home/agent/submission/submission.csv, and calls answer to submit. Reward is (score - bad_baseline) / (max_baseline - bad_baseline), where baselines are competition-specific. It is 0 at the bad baseline and 1 at the good one. It is floored at -1, because on tasks whose two baselines are close a slightly worse model would otherwise score far below -1; it is not capped above — an agent that beats the good baseline scores above 1, and scores are not comparable across competitions.

Only answer produces a reward; every other tool returns a blank one. A grading run that completes but yields no score (an invalid submission) returns a blank reward rather than a penalty and does not end the episode, so the agent can fix the file and resubmit; and a grading run that fails outright is retried three times and then raised, ending the rollout — no fabricated score in either case.

Data

Competition datasets (CSV files) are mounted read-only from cloud storage. Each competition includes:

• train.csv - training data with labels
• test.csv - test data for predictions
• sample_submission.csv - required submission format
• description.md - competition objectives and evaluation metric

Tools

• bash - Execute shell commands in the sandbox (600s per command)
• answer - Submit the final submission CSV for grading; a valid submission ends the episode, an invalid one is not scored and can be resubmitted
• read / write / edit / multi_edit - Read and modify files
• glob / grep / ls - Find files and search their contents
• todo_write - Track a task list

All but bash and answer come from the SDK's CLIToolset.

Time Horizon

MLEng is a multi-turn environment. Agents read competition descriptions, explore data, develop models, train and iterate, validate submissions, and submit for final scoring.

Environment Difficulty

[Put environment difficulty here]

Other Environment Requirements

There are no further environment requirements; MLEng works out of the box with the OpenReward endpoint without any external API keys.

Safety

Agents in MLEng develop ML models in a sandboxed environment. The environment does not present direct safety risks.

Citation

@software{mleng2025,
  title={MLEng: Machine Learning Engineering Environment for OpenReward},
  author={{General Reasoning Inc. Team}},
  year={2026},
  url={https://www.openreward.ai/GeneralReasoning/MLE-Eng-15}
}