Role Overview
Sciforium is seeking a highly technical and visionary LLM Dataset Engineer to lead the strategy, creation, and curation of the massive datasets that power our foundation models. We believe that in the era of LLMs, data is the primary competitive advantage. In this role, you will own the end-to-end data lifecycle—from raw web-scale crawling to the fine-grained human-alignment datasets that define model behavior.
This position is ideal for a scientist who views data as a high-scale engineering challenge and an analytical puzzle. You will not just "provide" data; you will design the taxonomies, filtering heuristics, and post-training pipelines that ensure our models are world-class in reasoning, safety, and multimodal understanding.
Key Responsibilities
- Foundation Dataset Strategy: Own the end-to-end creation of pre-training datasets for LLMs. This includes defining the mix of web data, code, books, and technical papers to optimize for downstream model performance.
- Petabyte-Scale Curation: Design and implement sophisticated pipelines for data cleaning, exact/fuzzy deduplication, and high-quality signal extraction from petabytes of raw, unstructured data.
- Post-Training & Alignment Data: Lead the development of high-quality post-training datasets, including Supervised Fine-Tuning (SFT) instructions, multi-turn dialogues, and preference modeling data (RLHF/DPO).
- Multimodal Expansion: Drive the acquisition and processing of vision and video data, navigating the complexities of multimodal alignment, video compression, and temporal data consistency.
- High-Performance Engineering: Develop high-throughput data processing scripts using Python, leveraging multiprocessing and multithreading to handle massive-scale ingestion and transformation without bottlenecks.
- Data Profiling & Analysis: Conduct deep-dive statistical analysis on training corpora to identify biases, gaps in knowledge, and quality regressions, ensuring the "diet" of the model is mathematically balanced.
- Synthetic Data Generation: (Added Value) Design pipelines to generate high-reasoning synthetic data to augment gaps in natural datasets, utilizing existing models for data labeling and refinement.
Must-Haves
- 5+ years of industry experience in Data Science or Machine Learning, with a proven track record of building and managing datasets for foundation models.
- Deep Proficiency in Python: Expert-level skills with a focus on high-performance code, including multiprocessing, multithreading, and efficient memory management for large-scale data tasks.
- Petabyte-Scale Experience: Demonstrated experience working with petabyte-scale datasets that have been directly used to train production-grade LLMs or Large Vision Models.
- Dataset Reconstruction: Experience building massive LLM training sets from scratch, including raw web crawls (e.g., Common Crawl) and specialized domain data.
- Post-Training Expertise: Hands-on experience building datasets for RLHF, DPO, and multi-turn instruction following, including the management of human-labeling workflows and quality gold-sets.
- Data Tooling: Mastery of data-at-scale frameworks such as Spark, Ray, or high-performance data-loading formats (e.g., WebDataset, Parquet).
Nice-to-Haves
- Computer Vision (CV) Curation: Experience building large-scale image or video datasets from scratch (e.g., LAION-style pipelines).
- Multimodal Crawling: Familiarity with large-scale crawling of multimodal data and the associated challenges of video processing, codecs, and compression.
- Taxonomy Design: Experience in designing complex labeling schemas for reasoning, coding, and mathematical benchmarks.
- Research Background: A Master’s or PhD in a quantitative field with a focus on data-centric AI or information retrieval.