Sabrina Klajda

Key Skills

Software

Appen

Clickworker

CrowdSource

CVAT

Data Annotation Tech

Labelbox

Mindrift

OneForma

Scale AI

Telus

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Text

Video

Top Label Types

Evaluation Rating

Prompt Response Writing SFT

RLHF

Text Generation

Translation Localization

Freelancer Overview

Applied Reinforcement Learning from Human Feedback (RLHF) to evaluate and rank AI-generated output Developed and refined rubrics to guide output evaluation Conducted red-teaming and safety assessments, identifying potentially harmful, biased, or non-compliant AI behavior Performed stereotype and bias analysis to support ethical AI development Created "gold standard" answers and corrected AI outputs to train high-quality generative models Output comparison tasks across a variety of dimensions (e.g. Instruction Following, Accuracy/Factuality, Localization, Writing Quality, etc.) Annotated large-scale datasets across multiple domains (audio, text) Participated in audio transcription and training tasks, contributing to multilingual speech recognition systems Bilingual: German and English Including STEM-focused projects Worked as both Contributor and Trusted Reviewer

Labeling Experience

Training models to respond safely to unsafe prompts

Data Annotation TechTextRed Teaming

Prompt (with picture) would be provided, two model outputs. Judging model responses on whether they respond appropriately to the harm level of the prompt. (e.g. very harmful prompts should be met with a refusal, prompts with sensitive subject matters should receive a disclaimer)

Prompt (with picture) would be provided, two model outputs. Judging model responses on whether they respond appropriately to the harm level of the prompt. (e.g. very harmful prompts should be met with a refusal, prompts with sensitive subject matters should receive a disclaimer)

2025

Rubrics creation

Data Annotation TechTextEvaluation Rating

Creating a prompt, receiving three different model outputs. Choosing a preferred model output, editing model output to be perfect, and then creating extensive rubric criteria to evaluate the model responses by, choosing which response performed the best.

Creating a prompt, receiving three different model outputs. Choosing a preferred model output, editing model output to be perfect, and then creating extensive rubric criteria to evaluate the model responses by, choosing which response performed the best.

2025

Prompt and Response Writing (Single-Turn/Multi-Turn)

Scale AITextPrompt Response Writing SFT

Providing a complex prompt, and then editing the response to be "gold standard". Both in Single-Turn and Multi-Turn. Also writing "gold standard" responses from scratch to feed the model training data. Evaluation of the response on several dimensions (IF, accuracy, localization, etc.)

Providing a complex prompt, and then editing the response to be "gold standard". Both in Single-Turn and Multi-Turn. Also writing "gold standard" responses from scratch to feed the model training data. Evaluation of the response on several dimensions (IF, accuracy, localization, etc.)

2024

Bilingual Output Evaluation

Scale AITextRLHF

Creation of a complex prompt to evaluate the performance of two different models side-by-side on various dimensions, such as Instruction Following, Accuracy/Factuality, Localization and Writing Quality.

Creation of a complex prompt to evaluate the performance of two different models side-by-side on various dimensions, such as Instruction Following, Accuracy/Factuality, Localization and Writing Quality.

2024

Education

Work History