Research-Oriented Internship
Internship · Year 1 · Period 3 · 20 ECTS · mandatory
Contents
| Module | Research-Oriented Internship |
| Period | M1 · Period 3 |
| ECTS | 20 |
| Status | Mandatory · research internship |
Stakes (Extrapolating)
Official unit for 2026–2027: INT_52402_EP — 16-week research-oriented internship (academic or in a public/private company), listed in LLGA 1 – S2 and taught in period 3; 20 ECTS per the LLGA wiki. The rest of this page is our own framing. The M1 research internship is the first major test outside the classroom: state a question, read the literature, experiment, and report. Empirically, this is where the Graph×LLM track becomes a topic rather than a pile of modules.
It is not yet the end-of-studies internship (24 ECTS in M2): the accent is research-oriented — academic lab, industrial R&D, or hybrids — with a report / poster / reproducible code deliverable depending on supervision.
Place in the 2-year path
Upstream: P1–P2. Here: P3 immersion. Downstream: M2 specialization then end-of-studies internship. Many M2 topics read better after touching one paper end-to-end in M1.
See After the master on the home page for the roles and problem spaces the track opens onto; they help when choosing the direction of a topic.
From your Year 1 project to an internship proposal
The Year 1 end-of-year projects are a dress rehearsal for this internship. They run until about two weeks after the end of Period 2, before it starts, and they follow the rules of this page: one question, one metric and one baseline per work package, a weekly log, at least 3 seeds and a reproducible repository. Their last deliverable is a one-page open question for the research internship (deliverables). That page is the seed of your topic.
Turn it into a two-page proposal before the internship begins. Reuse the project’s open question, its metric and its retrained baseline, so that the internship starts from a number you already trust. Then state three things: the gain you expect over the project result, what the host brings (data, compute, expertise), and the Year 1 chapters the work relies on. The host will reshape the question; the proposal still shows that you can scope one.
Each project prepares a family of internship topics, which a Year 2 Transverse project can then continue. The table repeats the continuation table of the projects page.
| Year 1 project | Internship topic family it prepares | Year 2 project |
|---|---|---|
| Y1.1 KG-QA router | relation extraction, KG quality, retrieval evaluation | Y2.1 |
| Y1.2 fleet dispatch | spatio-temporal forecasting, RL for operations | Y2.4 |
| Y1.3 news recommendation | recommender systems, bandits, privacy or fairness | Y2.3, Y2.5 |
| Y1.4 molecules from words | molecular ML, generative chemistry | Y2.2 |
| Y1.5 scene-graph QA | perception, neuro-symbolic reasoning, efficient inference | Y2.3 |
Industry and research links
Research: experimental method, baselines, ablations, writing. Industry R&D: the same evidence bar with data/prod constraints. Helps choose for M2: doctoral research, ML engineering, or CIFRE frontier.
How to approach it
- One question, one metric, one baseline — before stacking models.
- Weekly log: hypothesis → trial → result → next question.
- Keep an explicit link to at least one Year 1 course you took, for example Graph ML ch. 10 (GraphRAG and the hybrid), Text Mining & NLP ch. 7 (transformers, a minimal RAG pipeline), Reinforcement Learning, Multimodal Generative AI or Optimal Transport. The Large Language Models course comes later, in Year 2, Period 1.
- Evidence bar: retrain the baseline in your own pipeline, use at least 3 seeds, compare with a paired test, ablate each new component and touch the test set once. This is the protocol of Machine Learning 4.9; the shared evaluation kit of the end-of-year projects gathers the other tools.
A 16-week milestone template
The template below is a proposal of these notes. Adapt it to the host’s calendar; the supervisor sets the real milestones.
| Weeks | Milestone | Evidence at the end |
|---|---|---|
| W1–2 | Two-page proposal (question, metric, baseline, course anchor) and a literature map of 8–10 annotated papers | proposal agreed with the supervisor |
| W3–4 | Baseline reproduced in your own pipeline, on the host’s data or a public proxy | baseline number within the reported variance, or the gap explained |
| W5–10 | Experiments, one hypothesis at a time, recorded in the weekly log | log entries and a results table over seeds |
| W11–12 | Ablations: remove each new component in turn | ablation table with standard deviations |
| W13–14 | Writing: full draft, figures regenerated by scripts | draft sent to the supervisor |
| W15–16 | Poster or defence, repository freeze | tagged release, poster or slides |
Reproducibility checklist
- Seeds: at least 3 per trained model, reported as mean and standard deviation; seeding as in Deep Learning 8.4.
- Configs: every run starts from a versioned configuration file, saved with the commit hash next to its result.
- Lockfile: exact package versions (
uv.lockor a pinned requirements file) and, when possible, a Dockerfile, as in the repository template of the CS refresher, 3.9. - Data versions: a hash or a dated snapshot of each dataset, and the script that builds the splits.
- One command: each reported number is regenerated from the README with one command, and a peer has tried it.
Report skeleton
- Abstract: the question and the answer, with the headline number and its spread.
- Introduction: why the question matters, and the contributions in two or three bullets.
- Related work: the literature map, condensed to the papers you build on or compare with.
- Method: what is new, with the baseline it extends.
- Experimental setup: data and its version, splits, metric, baseline, seeds, compute budget.
- Results: mean and standard deviation over seeds, paired tests against the baseline.
- Ablations and failure analysis: what each component brings, and where the method fails.
- Conclusion and open question: what is settled, and the next question, which can seed the Year 2 work.
- Appendix: the commands, configurations and data hashes behind each number.
Extrapolating Framing; internship rules and grading = programme / host lab.