Internship – Fine-tuning Large Language Models for Contextualised Outputs – Data Integration Department – WOAH HQ – Paris, France

Internship – Fine-tuning Large Language Models for Contextualised Outputs – Data Integration Department

Context

The World Organisation for Animal Health (WOAH – founded as OIE) is a leading intergovernmental organisation representing 183 Members worldwide. Through its activities, WOAH makes a decisive contribution to improving animal health, protecting animal welfare and strengthening Veterinary Services. The Organisation provides transparent information on world's animal health situation, and promotes international standards, particularly in terms of the safety of trade in live animals and animal products. More information can be found on
WOAH's website
.

Joining WOAH means taking part in the development of one of the leading international organisations, recognised and associated with other multilateral institutions, in the field of worldwide health. It means helping to build a global approach to health, combining animal and human health in a "One Health" approach. It means joining teams motivated by the impact of their actions, the sense of their collective commitment and their recognised professionalism in their respective fields of expertise.

WOAH's headquarters are based in Paris. The Organisation is present on every continent through 13 Regional or Sub-regional Representations. WOAH has 250 staff members, two-thirds of whom are based at headquarters.
This position is located in Paris, France

The World Organisation for Animal Health (WOAH) depends on accurate, consistent and timely textual information for epidemic intelligence, rapid risk assessment and Member support. Many WOAH workflows — translation of technical guidance, synthesis of surveillance reports, classification of signals, and generation of standardised summaries — depend on consistent use of terminology (glossaries), examples of prior translations, and adherence to editorial/guideline rules (e.g., English usage, style for veterinary terminology).

Large Language Models (LLMs) offer powerful capabilities for translating text, but off-the-shelf models can be inconsistent when required to follow domain glossaries, replicate specific translation patterns, or respect detailed editorial guidelines. Fine-tuning and model adaptation workflows can substantially improve consistency and reliability by incorporating: (i) official glossaries/terminology lists; (ii) curated past examples of high-quality translations and editorial corrections; and (iii) explicit guidelines on style and register. These adaptations may be performed via cloud API fine-tuning (OpenAI, Google Gemini) or by local/lightweight adaptation methods (LoRA/QLoRA) where appropriate. The WOAH Data Integration Department (DID) needs a reproducible, documented, and easily maintained set of workflows and artifacts that ensure LLM outputs are aligned with WOAH standards.

Job Description

Positioning and reporting

The intern will be placed within ISS Directorate and report directly to the Head of the Data Integration Department, with functional guidance from WOAH scientists involved in AI tools uptake. The intern will work closely with translation/editing staff and the team maintaining WOAH glossaries and style guides.

Intership purpose – Master thesis project

Design, implement and evaluate reproducible fine-tuning and adaptation workflows to incorporate domain context (glossaries, prior translation examples, and English/style guidelines) into LLM outputs that support WOAH's translation, summarisation and RRA text-generation tasks. The project should prioritize practical, maintainable pipelines that DID staff can run or orchestrate from R and/or command-line tools, and that clearly document trade-offs between cloud API tuning and local/lightweight adaptation.

Missions and activities

Requirements & data collection

• Work with WOAH editors/linguists to gather glossaries, style guides, and a representative set of prior translations and editorial corrections (anonymised where necessary).

• Define the target tasks and acceptance criteria (terminology accuracy, translation fidelity, adherence to style rules, reduction in manual editing time).

Design adaptation strategies

• Review and compare adaptation approaches: cloud API fine-tuning (OpenAI GPT-5, Google Gemini Flash-2.5), prompting strategies (system messages, few-shot examples), and local adaptation (LoRA/QLoRA, adapters, prompt engineering).

• Propose two to three candidate strategies for different operational contexts (full cloud fine-tune, lightweight local adaptation, hybrid workflow).

Data preparation and tooling

• Prepare training/validation datasets in the formats required by the selected fine-tuning platforms (JSONL for OpenAI, JSON for Gemini, LoRA datasets for local methods).

• Implement reproducible scripts (R and/or Python) to perform tokenisation estimates, convert formats, and upload or launch fine-tuning jobs.

Implementation and experiments

• Conduct controlled fine-tuning/adaptation experiments on at least two approaches (e.g., cloud fine-tuning for OpenAI or Gemini; LoRA adaptation on an open model locally).

• Evaluate using quantitative metrics (terminology accuracy, BLEU/ChrF or similar metrics, editorial guideline adherence) and qualitative review by WOAH editors.

Cost, compute and operational assessment

• Document expected cloud fine-tuning costs (OpenAI, Gemini) based on WOAH datasets (100–1000 pages).

• Analyse operational implications (compute location, data handling, iteration speed, maintainability).

Documentation, reproducibility and handover

• Deliver clear step-by-step pipelines, annotated code, and operational instructions for DID staff (including R scripts to call cloud APIs).

• Produce final technical report, reproducible notebooks, and an executive summary for non-technical stakeholders.

Design and implementation of a simple user interface (GUI)

• Develop a prototype graphical interface (in
Dash
or
RShiny
) to allow WOAH staff to:

– upload small samples of text for testing;

– select between models (baseline, fine-tuned, LoRA-adapted);

– provide glossary terms or contextual examples;

– view side-by-side outputs (baseline vs tuned model) for editorial evaluation.

• Ensure that the GUI is lightweight, easy to deploy locally, and documented so that DID technical staff can maintain or extend it.

Requirements

Qualifications and Experience

Minimum Required qualifications

– Master level (or equivalent) in data science. Alternatively in veterinary sciences, life science, international affairs, public health, agricultural science with formal training or experience in data science.

Expected Skills

Technical skills:

– R or python advanced level

– Familiarity with LLM fine tuning theory and methods would be a plus

– Analysis and problem solving.

– Good working knowledge of Microsoft Office, such as Excel, Word and PowerPoint.

– Excellent knowledge of English, both spoken and written.

– Basic knowledge of epidemiology and epidemic intelligence principles

Interpersonal skills:

– Ability to work in an agile way and deliver at pace… organisation skills and ability to meet deadlines.

– Team player who can integrate well into the department and is willing to commit to supporting the development of Observatory outputs.

Additional Information

Working conditions

The post is a full-time position based at the WOAH Headquarters in Paris. It requires long hours in a seated position at a computer.

Salary:
800 euros/month

Duration:
4 to 6 months – Starting date flexible from December 2025 to October 2026

General Information

WOAH places high value on a multicultural and positive work environment. WOAH is an equal opportunity employer and welcomes all qualified candidates, irrespective of their origin, gender, opinions or beliefs.

If you are interested in the position, please complete your application online at the latest by January 5th, 2026.

APPLY HERE