Skip to content

Prove your machine learning skills

Many people have trained a model on a clean tutorial dataset. Fewer can explain why a model that scored well offline fails in use. Lemma records that judgement through live reviews, assessments and projects you connect.

How machine learning is assessed

What machine learning competence looks like

A competent ML practitioner asks whether a model is needed before building one, and sets a simple baseline first. They split data so nothing leaks from the future or from duplicate records, choose a metric that matches the cost of each kind of error and look closely at where the model is wrong.

They debug models methodically. When performance drops, they check the data pipeline, label quality and distribution shift before tuning hyperparameters. They can explain the bias and variance trade-off in their own model, document its limits and plan how it will be monitored once real inputs start arriving.

Machine learning session
Live · 41:07

Reviewing a model that underperforms

60 min · Learning · with Zara Q.

train.py Daniel O. is sharing
12
13
14
15
16
17
18
19

Zara Q. rated the session

Practitioner · after the call

Clear reasoning on problem framing and baselines. Talked through trade-offs first.

Added to Machine learningAdded to record

Problem framing and baselines

Deciding whether ML fits, defining the target and beating a simple baseline before adding complexity.

Data splits and leakage

Train, validation and test splits that respect time and groups, and spotting leakage early.

Model evaluation

Choosing metrics that match real error costs, calibration and error analysis by segment.

Feature engineering

Building informative features, handling categorical and missing data, and avoiding target leakage.

Training and tuning

Regularisation, cross-validation and tuning with a budget, without overfitting to the validation set.

Deployment and monitoring

Serving models, tracking drift and deciding when a model needs retraining or retiring.

What counts as evidence of machine learning skills?

In machine learning, a polished result can hide a shaky method. Lemma records the evidence behind the method: sessions where someone followed your reasoning, an adaptive test and the repositories you link, notebooks included. Each source is listed below. See the evidence methodology for how each source is weighed.

Skills Passport
Signed
64

Machine learning

Mei L. · SCI recomputed 9 Oct

Evidence · 4 entries

  1. Reviewing a model that underperforms

    Live session · 9 Oct

    Rated 5.0

  2. ML debug challenges

    Adaptive assessment · 2 Oct

    Level 7.2

  3. An ML project on GitHub

    GitHub connector · 24 Sep

    Verified

  4. Teaching an ML concept

    Live session · 11 Sep

    Taught · 1.5x
Signature valid · checked today Share link
Live session

Reviewing a model that underperforms

A practitioner presents a model with a hidden problem, such as leakage or a bad split. In a live session you diagnose it and explain the fix, and they rate your reasoning.

Adaptive assessment

ML debug challenges

Adaptive scenario questions cover evaluation, overfitting and data issues. At higher difficulty you fix broken training or evaluation code, and the fix and explanation are scored against a rubric.

GitHub connector

An ML project on GitHub

Connect a repository with a training pipeline, evaluation and write-up. Lemma checks it exists and that your account has access, and re-checks it periodically.

Live session

Teaching an ML concept

Explaining cross-validation or evaluation metrics to another learner with a worked example shows depth. Teaching sessions count for more than learning ones in your record.

How is a machine learning skills test run on Lemma?

The machine learning test asks 15 to 25 questions and moves up or down a level with each answer, focusing on judgement calls such as leakage, skew and bad joins. It feeds your machine learning SCI together with sessions and linked work, and the score weighs how fresh that evidence is.

How the SCI score is calculatedHow adaptive skills assessments workPractise machine learning in live skill swap sessionsWhich evidence types countWhat your Skills Passport shows
Machine learning assessment
0:42
Question 11Stops when your level holds
Difficulty7/10

Validation accuracy is 99% but the model fails in production. One feature is filled in after the outcome. What is wrong?

  1. AThe model is underfitting
  2. BThe learning rate is too high
  3. CTarget leakage from that feature
  4. DThe test set is too large
Locks on submit

  1. Practise core concepts

    Work through AI tutor exercises on evaluation, regularisation and feature handling, with explanations in your own words. The tutor adapts to you. Drill results add only a small, capped amount to your score.

  2. Review models with a practitioner

    In live sessions, present a model or diagnose one together. A practitioner probes your choices of split, metric and features, then rates the session as evidence for machine learning.

  3. Pass the ML assessment

    Scenario questions on evaluation and data problems adjust as you answer. Higher difficulty brings code-level debugging. Failed attempts slightly lower your proof score, so preparation pays off.

Machine learning skills: common questions

How do I prove machine learning skills without industry experience?

Build evidence that shows judgement. Diagnose and present models in live sessions with practitioners, pass the adaptive assessment with its debug challenges, and connect a project that includes honest evaluation and error analysis. That record shows how you work, which is what a hiring team is trying to learn.

What does a machine learning skills assessment cover?

Lemma's ML assessment focuses on problems that break real models: leakage, poor splits, mismatched metrics, overfitting and distribution shift. Questions are scenario-based and adapt to your level. At higher difficulty you fix broken training or evaluation code and explain what was wrong.

Is a Kaggle-style project enough for an ML portfolio?

Competition work shows you can optimise a metric on a fixed dataset. Teams also want to see problem framing, data cleaning, baselines and how you would monitor a model in use. Add those to your project write-up, connect the repository and discuss it in a session for a fuller picture.

Do I need deep learning experience for this skill?

No. The skill covers machine learning broadly, and most of it, including evaluation, splits and feature work, applies to classical models and neural networks alike. If deep learning is your focus, bring it into your sessions and projects so the evidence reflects it.

Skills related to machine learning

Browse the skills directory

Python

Show how you write, debug and structure Python, not just that you once finished a course.

Evidence for Python

Data analysis

Evidence of how you frame questions, clean data and reach conclusions a decision-maker can rely on.

Evidence for data analysis

Data engineering

Evidence that your pipelines are reliable, testable and recoverable, reviewed by practitioners in live sessions.

Evidence for data engineering

Cloud architecture

Show how you design cloud systems that scale, fail gracefully and stay affordable.

Evidence for cloud architecture

Start your machine learning evidence record.

Lemma is in early access. Request an invite and we will email you when your place is ready.