Language models,
done hands-on.
Workshops at Xetun Bavori put you inside the mechanics — tokenisation, embeddings, fine-tuning, evaluation. Each session is structured around a task you can finish, not a lecture you sit through.
Where the curriculum comes from
NLP moves fast. Instructor notes from 2022 no longer reflect how practitioners actually build and deploy text pipelines.
Workshop content at Xetun Bavori is reviewed against recent arXiv publications, Hugging Face model releases, and practitioner threads from communities like r/MachineLearning. When spaCy updates its pipeline API or a new tokenizer architecture gains traction, the affected exercises are rewritten before the next cohort starts.
The platform was built in 2014 to solve one problem: the gap between what classrooms taught and what engineers actually needed on the job. That gap is still the thing this work is aimed at.
What participation looks like, practically
Per week, across asynchronous exercises and two live sessions. Sessions are recorded for participants in different time zones.
Each workshop track runs six weeks. Assignments are sequential — each one builds on the output from the previous task.
Prerequisites differ by track. Most require comfort with Python and at least some exposure to statistics. No prior NLP knowledge is assumed for entry-level tracks. More advanced tracks specify what you should already know before enrolling.
"The peer review step on the named-entity task was the part where I actually started noticing my own assumptions about the data."
— Oles Verbytsky, text analytics engineerWho else is in the room
Participants arrive from different positions: some are software engineers adding NLP to their skill set, others are linguists moving toward applied computational work, a smaller number are researchers validating practical methods.
Cohort size is kept deliberately small — between 14 and 22 participants — so peer review and group debugging sessions stay substantive rather than performative.
A selection of current tracks
-
01
Text classification from scratch
Covers vectorisation strategies, model selection trade-offs, and evaluation metrics that go beyond accuracy — precision, recall, and where each matters in production.
-
02
Transformer architecture and fine-tuning
Participants fine-tune a pre-trained model on a domain-specific corpus. The exercises focus on data preparation decisions and how they affect downstream results.
-
03
Information extraction and entity recognition
Structured outputs from unstructured text — annotation schemes, relation extraction basics, and tools including spaCy and Prodigy for building a custom NER pipeline.
How the work here differs from self-paced video courses
Video courses transfer information. Watching someone else build a sentiment classifier does not produce the same knowledge as debugging one yourself — particularly when the data is messier than the example data.
Workshops here are structured around deliberate failure. Exercises are designed so that a first attempt will expose gaps. Feedback from instructors is specific to what you submitted, not to a generic answer key.
The difference is not ideological — it is structural. Cohort-based scheduling, fixed deadlines, and peer review create conditions where sustained effort is more likely. No format guarantees results; this one reduces the friction that usually stops self-directed learning before it produces durable skill.
Read more about the approachInstructors respond to your specific code and reasoning — not templated guidance.
Fixed start dates and shared deadlines keep progress social rather than isolated.
Exercises use datasets with real noise — missing labels, imbalanced classes, ambiguous annotations.
All materials and session recordings available globally without time zone restrictions.