Overview

Build data pipelines for AI model training.

Responsibilities

• Source and ingest large-scale datasets from Common Crawl, public web sources, open-source repositories and other data sources.
• Build scalable data processing pipelines for large unstructured corpora.
• Develop and optimise distributed data pipelines using tools such as Apache Spark, Ray and Dask.
• Extract and process content from web pages and documents using Trafilatura and related web/content extraction libraries.
• Design data cleaning, filtering, quality assessment and curation pipelines for LLM pretraining corpora.
• Develop heuristic-based filters to remove low-quality, boilerplate, spam, duplicated or otherwise unsuitable content.
• Implement document-level and corpus-level deduplication strategies.
• Identify and mitigate benchmark contamination and evaluation-data leakage in training corpora.
• Prepare and curate benchmark and evaluation datasets.
• Work with large-scale datasets stored and processed through GCS/GCP.
• Monitor data quality, pipeline performance, storage and processing efficiency.
• Collaborate with AI Researchers and ML Engineers to continuously improve corpus quality and training-data pipelines.

Qualifications

• 2-5 years of experience in Data Engineering, AI Data Engineering, ML Data Engineering or a related field.
• Strong Python programming skills.
• Strong understanding of data cleaning, filtering, validation and deduplication.
• Hands-on experience with exact and near-duplicate detection and deduplication.
• Familiarity with at least one distributed processing framework: Apache Spark, Ray or Dask.
• Understanding of web-scale data ingestion, such as Common Crawl or other large public datasets.
• Familiarity with GCS/GCP and cloud-based data processing.
• Strong problem-solving and ownership mindset.
• Familiarity with Hugging Face Datasets, DataTrove or NVIDIA NeMo Curator.
• Knowledge of MinHash, fuzzy or semantic deduplication.
• Research-level understanding of synthetic data generation techniques for model training and evaluation.
• Familiarity with Rust for high-performance data processing.