About the job
As a Research Scientist/Engineer focused on honesty within the Finetuning Alignment team, you'll spearhead the development of techniques to minimize hallucinations and enhance truthfulness in language models. Your work will focus on creating robust systems that are accurate and reflect their true levels of confidence across all domains, and that work to avoid being deceptive or misleading.
Responsibilities
Design and implement novel data curation pipelines to identify, verify, and filter training data for accuracy given the model’s knowledge
Develop specialized classifiers to detect potential hallucinations or miscalibrated claims made by the model
Create and maintain comprehensive honesty benchmarks and evaluation frameworks
Implement techniques to ground model outputs in verified information, such as search and retrieval-augmented generation (RAG) systems
Design and deploy human feedback collection specifically for identifying and correcting miscalibrated responses
Design and implement prompting pipelines to generate data that improves model accuracy and honesty
Develop and test novel RL environments that reward truthful outputs and penalize fabricated claims
Create tools to help human evaluators efficiently assess model outputs for accuracy
Qualifications
Minimum
Have an MS/PhD in Computer Science, ML, or related field
Possess strong programming skills in Python
Have industry experience with language model finetuning and classifier training
Show proficiency in experimental design and statistical analysis for measuring improvements in calibration and accuracy
Care about AI safety and the accuracy and honesty of both current and future AI systems
Have experience in data science or the creation and curation of datasets for finetuning LLMs
An understanding of various metrics of uncertainty, calibration, and truthfulness in model outputs
Preferred
Published work on hallucination prevention, factual grounding, or knowledge integration in language models
Experience with fact-grounding techniques
Background in developing confidence estimation or calibration methods for ML models
A track record of creating and maintaining factual knowledge bases
Familiarity with RLHF specifically applied to improving model truthfulness
Worked with crowd-sourcing platforms and human feedback collection systems
Experience developing evaluations of model accuracy or hallucinations