Internship — end-of-studies internship
Internship · Year 2 · Period 3 · 24 ECTS · mandatory
Contents
| Module | Internship — end-of-studies internship |
| Period | M2 · Period 3 |
| ECTS | 24 |
| Status | Mandatory · final internship |
Stakes (Extrapolating)
Extrapolated notes. The end-of-studies internship (24 ECTS) is the master’s capstone: broader ownership than in M1, often a professionalizing setting (company, applied lab, startup), and a sustainable deliverable. Empirically, this is where the Graph×LLM track turns into a job or a research direction: After the master, on the home page, lists the roles and problem spaces it opens onto.
Place in the 2-year path
Upstream: all of M1 including the research internship, M2 specialization, ethics seminar, transverse project. Here: long P3 execution. Typical difference vs M1 research internship: more product / team / ops integration, even when the topic stays R&D.
The Year 2 end-of-year projects that these notes propose for the Transverse project end just before this internship and prepare it. Their last deliverable is a one-page open question for the final internship, or a PhD / CIFRE proposal (deliverables). Their charter already holds a module map, and their trust work package and ethics memo rehearse the report’s risks, ethics and data section. When the host’s topic allows, start the internship from that open question and the baseline your project measured; the module map below names the project behind each topic family.
Industry and research links
Industry: productionization, impact measurement, multi-team collaboration. Research: possible PhD path (academic or CIFRE) if the internship produced a solid open question. P1/P2 management and M2 ethics help frame “should we deploy?” not only “can we run it?”.
How to approach it
- One-sentence goal + measurable success criteria by week 1–2.
- Explicitly link the topic to 2–3 chapters of these notes (path evidence), using the module map.
- Plan a risks / ethics / data section in the report — not a cosmetic appendix.
Module map: the chapters behind your topic
Each row is a topic family that final internships often meet. It lists the chapters of these notes that back it and the Year 2 project that rehearses it. In the report, cite the 2–3 chapters you actually used and say what each one brought.
| Topic family | Chapters that back it | Year 2 project |
|---|---|---|
| GraphRAG and knowledge-graph question answering | Graph ML ch. 10 (GraphRAG and the hybrid), Text Mining & NLP 6.4 (relation extraction), Large Language Models ch. 7 (RAG in depth, 7.4 GraphRAG) | Y2.1 |
| Recommender systems | Text Mining & NLP 8.5 (content-based and collaborative), Reinforcement Learning ch. 4 (bandits), Graph ML ch. 6 (link prediction) | Y2.3 |
| Fraud and financial crime | Advanced GNN ch. 6–7 (temporal graphs, robustness and privacy on graphs), Boosting to Foundation Models ch. 9 (honest benchmarking), Security and Robustness ch. 9 (attacks on graphs and LLMs) | Y2.5 |
| Molecules and materials | Graph ML ch. 9 (graph generation), Graph Generative Models ch. 7–8 (conditional generation, drugs and materials), Geometric Data ch. 7–8 (Euclidean equivariance, applications) | Y2.2 |
| Serving and operations | Large Language Models ch. 6 (inference), ML Ops and LLM Engineering (ch. 3 serving, ch. 5 monitoring) | Y2.4, Y2.3 |
| Trustworthy AI | Large Language Models ch. 9 (evaluation and safety), Privacy and Uncertainty Quantification (differential privacy, conformal prediction), Explainability, Security, Privacy of LLMs (attacks, red teaming) | the trust work package of every Year 2 project |
Report skeleton and rubric
The skeleton and the rubric are proposals of these notes; the programme and the host set the real format and grading. Draft sections 1–3 in the first two weeks, then check them against the final text.
- Goal and success criteria: the one-sentence goal and the 3–5 measurable criteria fixed in week 1–2, each marked met, partly met or missed.
- Starting point: the open question of your Year 2 project, or the host’s problem when it differs, and the baseline number you start from.
- Path evidence: your row of the module map and the 2–3 chapters reused, with what each one brought.
- System and method: what the host had, what you built, and the data with its provenance.
- Results: every headline number against the baseline, with a confidence interval over seeds or bootstrap resamples; impact or online metrics if the system shipped.
- Risks, ethics and data: licence and consent of the data, personal data, measured failure modes, misuse, and your answer to “should we deploy?”.
- Open question: what stays open after the internship; a solid one can seed a PhD or CIFRE proposal.
- Appendix: the commands, configurations and data versions behind each number.
| Criterion | Weight | What earns the marks |
|---|---|---|
| Goal and success criteria | 15 % | criteria stated early, measurable, and honestly assessed at the end |
| Path evidence | 10 % | named chapters, each with a concrete use in the work |
| Technical depth and correctness | 25 % | a sound method, a fair baseline, choices justified against alternatives |
| Results with confidence intervals | 20 % | intervals on every headline number, ablations, test data touched once |
| Risks, ethics and data | 15 % | a measured failure mode, data provenance, a reasoned deployment verdict |
| Open question | 5 % | precise, testable, and tied to the results |
| Communication and reproducibility | 10 % | clear writing and figures; each number regenerated from the appendix |
Extrapolating Framing; internship rules = programme and host organization.