Freelancer Overview
For most of the projects I've worked on as a Data Annotator and AI Content Evaluator, I reviewed and assessed AI-generated responses for relevance, accuracy, and overall quality, whilst identifying and flagging formatting issues, inconsistencies, biases, and factual errors. With these projects, I'd also detect subtle issues in logic, coherence, spelling, and grammar of the model's responses. With other projects, specifically focused on factuality, I have had to validate every single claim in a model's response, and locate and cite reliable and credible sources to support or contradict claims made in the model's response. Other projects would involve "stress testing" the models, by writing my own prompts (and sometimes my own system prompts too) and targeting the model's weaknesses, attempting to cause instruction following, factuality, or formatting issues.
I have good experience producing rubrics and answer keys for model responses, and editing a model's response into the "golden response" by removing all issues and writing the new, golden response in markdown. Additionally, I have worked on other more niche projects, which include a couple of projects involving image analysis, where I'd analyse AI generated images for anatomical mistakes, unintelligible text, and style inconsistencies, and I have worked on a couple of voice projects, where I'd contribute voice data to train models, helping improve their ability to recognise and understand different accents, speech patterns, and voices, in quiet or loud environments. Lastly, recently I have been working on a project where an AI model has access to various documents (emails, Google Docs, Google Slides, Google Sheets) and is often asked to summarise or rewrite the documents they have access to, and I'd have to verify the model's responses are fully grounded, truthful, helpful, and of a high quality.
