STEM OPT I-983 Formal Training Plans for Machine Learning Researchers

Why the I-983 matters more for ML roles

Form I-983 is the backbone of the 24-month STEM OPT extension. It is not a job description and it is not an offer letter — it is a formal training plan that explains how the practical experience relates directly to the student's qualifying STEM degree. For machine learning researchers this bar is higher than it looks, because reviewers frequently see plans that describe generic software engineering work rather than research training tied to statistics, optimisation or computational science coursework.

At DeltaDex Technologies we treat the I-983 as an engineering document. The same discipline that produces a reproducible training pipeline produces a defensible training plan: explicit objectives, measurable outcomes, named owners and a review cadence that generates evidence over time.

Writing objectives a reviewer can verify

Objectives should be written so an outside reader can tell what the trainee will be able to do at the end of each phase that they could not do at the start. Vague statements such as "gain exposure to AI" are the most common weakness. Strong objectives name the technique, the artefact and the standard of performance.

A useful pattern is: technique to be learned, applied to a defined problem, measured by a specific evaluation. For example, learning experiment tracking and ablation methodology, applied to a retrieval-augmented generation component, measured by the ability to independently design and defend an ablation study.

• Tie each objective to a course or competency from the qualifying degree.
• State the artefact produced: an evaluation harness, a model card, a benchmark report.
• Define the standard of independence expected by the end of the phase.
• Sequence objectives across the first and second twelve-month periods.

Supervision and the goals of the training relationship

Reviewers look for a real supervision structure rather than a nominal manager. Name the supervising researcher, describe their qualifications, and describe the mechanics of oversight: paper reading groups, design review of experiment plans, code review of pipeline changes, and one-to-one feedback sessions on research methodology.

The plan should also state how the employer's resources make the training possible. For ML researchers this typically includes access to labelled datasets under governed conditions, compute allocations for training runs, an experiment tracking system, and a registry that captures model lineage. These specifics distinguish a training environment from ordinary production work.

Evaluations and audit readiness

The self-evaluation at twelve months and the final evaluation are the parts most often completed late. Build them into the calendar the day the plan is signed, and keep the supporting evidence — experiment logs, review notes, published internal reports — alongside the signed form so that a site visit or a request for evidence can be answered from a single folder.

Material changes also matter. A change of supervisor, a substantial change in the trainee's duties, or a reduction in hours should trigger an amended plan rather than an informal note. Documented amendments are far easier to defend than a plan that silently drifted away from what was signed.