What the Project Milestone Reviews
This rubric looks at whether your milestone clearly explains the problem, data, baseline, model, early results, and plan for the remaining work. It also checks that your results are interpretable, your data process avoids avoidable problems, and your writing makes clear what the evidence does and does not show.
What We Check
Whether the milestone restates the problem as a question with a stated input and output, gives a sourced reason it matters, and uses the same target the data and results sections use.
Whether the milestone names the data and its origin, the preprocessing, the split with its rule and its unit, and the shape of what the model sees.
Whether the milestone names a baseline, describes the current model with how it was trained, credits the code it started from, and gives a reason for each model choice.
Whether the milestone reports at least one result with the metric defined, the model and split it came from, and something to compare it against.
Whether the milestone says what the numbers or the failed runs mean: whether they were expected and why, what failed or surprised, what caused it, and what the evidence cannot yet show.
Whether the remaining work is named step by step, tied to what the results showed, ordered or timed, and guarded by a fallback.
Mistakes We Flag
- Interest without a research reason
- Saying you chose a topic because you like it does not explain the gap, need, limitation, or problem the project addresses.
- Unclear data split or data leakage
- A split needs to say how records were assigned. Scaling or augmentation before splitting, or putting related records on both sides of a split, can make results less trustworthy.
- No meaningful comparison
- Naming models is not enough. Include a baseline and explain why each model choice is useful for the question you are testing.
- Results without context
- A number needs the metric, the model and data split it came from, and a comparison such as a baseline or class balance. Training results alone do not show how the model performs on held-out data.
- Vague next steps
- Plans such as “improve the model” do not show what you will do next. Name specific runs or changes, put them in an order or timeline, and connect them to what you learned so far.
How this rubric reads your essay
- Built from published reader guidance — the sources are listed below.
- Scored per section against defined criteria, not general impressions.
- Calibrated for consistent scoring, so a better draft shows up in the score.
References
A milestone should show the data, baseline model, and evaluation metric are set up with preliminary results compared against baselines, and should say whether the results were expected and what that means for next steps; code not written by the student must be made clear with a reference.
The course proposal questions the Inspirit template mirrors — the problem and why it is interesting, the challenges, the dataset and how it will be collected, the method, and how results will be evaluated, at 300–500 words — and the milestone's ask for a model description and a training strategy such as the loss function.
Explicit input and output, at least five related-work references, how many training, validation, and test examples and what preprocessing was done, a citation for the dataset's source, primary metrics explained before results, both quantitative and qualitative results, examples of failures, and whether the model overfit.
The milestone experiment need not be successful; if it did not work as expected, the writer should briefly discuss why.
Milestone rubric: a precisely stated, well-scoped problem; whether risks are highlighted; preliminary results presented and interpreted, or a concrete plan with a realistic timeline.
Procedures must include the methods for data collection and, when applicable, the source of data used, and must delineate what the student will do and what will be done by the mentor.
Every image, graph, table, and chart must be cited; no library research beyond the short introduction; research proposals and incomplete investigations are not eligible; work presented in the student's own words with disclosure of help received.
Judging criteria include the statement of the research problem, logical conclusions relevant to it, whether students recognize their contribution to the field, skill in communicating results, and references stated; the abstract must include the hypothesis and conclusions; presenters should avoid jargon or explain specialized terms.
The taxonomy of data leakage: no test set, preprocessing on training and test data together, temporal leakage, and records from the same people or units on both sides of the split.
Meaningful and naive baselines, accuracy on imbalanced classes (a classifier that always outputs the larger class scores 90% while being useless), sequential overfitting when tuning on the test set, and not generalizing beyond the data.
Train/validation/test split details, all preprocessing steps, hyperparameter ranges and selection method, the exact number of runs, and a clear definition of the reported measure.
Frequently Asked Questions
Do my preliminary results need to be successful?
No. A milestone can include a failed or unfinished experiment if you say what happened, give the likely cause when known, and explain what you will do next.
What should I include about my dataset?
Name the data and where it originally came from, describe preprocessing, explain the training, validation, and test split or cross-validation, and say how records were assigned to each part.
What makes a result useful to report?
Report a numerical metric from a run, identify the model and split, and compare it with a baseline or another meaningful reference. Then explain whether the result was expected and what it means.
Do I need to disclose code or help that I did not create myself?
Yes. Make clear when you started from existing code, a library model, pretrained weights, a tutorial, or a mentor’s implementation, and identify the source clearly.
Get your draft scored across all 6 dimension with specific, actionable feedback. Two free reviews per day — no credit card required.