Posted 1h ago
Loading this job…
NLP Intern - Hugging Face Transformers
Posted 2h ago0 views
What you will build
Job descriptions arrive on MailerMen as free text, and half of them bury the actual skills in a wall of paragraphs. In three months you will build the extraction layer that turns that text into structured skill tags, a seniority label and a clean summary line.
The stack
You start from a Hugging Face baseline, fine-tune a small encoder on roughly 12,000 hand-labelled postings, then compare it against a prompt-only approach so we know what the fine-tune is actually buying. Expect to spend real time on the label set itself: deciding whether React and React.js collapse into one tag, and what to do with the posting that lists nine languages and means none of them.
How we work
You pair with the platform engineer who owns ingestion, demo on Thursdays, and write a short teardown of every experiment that failed. The internship is fully remote inside India with one morning video sync each week. Everything else happens in pull requests and a shared results doc. You finish with one model running against live postings and an evaluation report you can show anyone who asks what you did.
Responsibilities
- Fine-tune a small transformer encoder on labelled job postings to extract skills, seniority and role family
- Curate the annotation guideline, including rules for collapsing near-duplicate skill names
- Benchmark the fine-tuned model against a prompt-only extraction baseline on a shared held-out set
- Build the inference script that runs extraction over newly ingested postings every night
- Investigate failure cases where a posting lists technologies the role does not really use
- Demo the week's experiment to the platform team each Thursday
- Write a short teardown for each experiment that did not reach production
Requirements
- Final-year student or recent graduate with coursework or self-study in natural language processing
- Working Python, plus familiarity with the Hugging Face transformers library or an equivalent framework
- Understands tokenisation, embeddings and the difference between fine-tuning and prompting
- Can explain precision, recall and F1 without looking them up
- Reliable internet and availability for one morning video sync each week
Skills
Benefits
- Paid monthly stipend with fully remote working anywhere in India
- Your extraction model runs against live postings rather than a sandbox copy
- Weekly pairing time with the platform engineer who owns the ingestion pipeline
- Certificate of completion plus a written recommendation on request