Darragh Byrne

Key Skills

Software

Data Annotation Tech

Internal/Proprietary Tooling

Top Subject Matter

No subject matter listed

Top Data Types

Document

Image

Text

Top Label Types

Action Recognition

Classification

Data Collection

Evaluation Rating

Fine Tuning

Object Detection

Prompt Response Writing SFT

Question Answering

Relationship

Text Generation

Text Summarization

Tracking

Translation Localization

Freelancer Overview

I am an experienced AI Trainer and LLM Evaluator with over 2.5 years of hands-on work in high-accuracy data annotation, specializing in factuality, citation integrity, and complex instruction adherence for large language models. My background includes evaluating AI outputs for subtle hallucinations, misleading claims, and dialect alignment—particularly Hiberno-English—using custom dashboards and rubric-based QA frameworks. I have consistently passed quality audits and contributed to advanced projects requiring nuanced judgment, such as multi-turn conversations, long-instruction prompts, and citation validation. I excel at providing detailed written feedback, peer reviewing annotation work, and adapting quickly to evolving guidelines, all while working independently in fully remote environments. My analytical, evidence-based approach ensures high standards of model honesty, reliability, and cultural accuracy.

Labeling Experience

Teaching AI To Be More Honest - Long Instructions

Internal Proprietary ToolingTextQuestion AnsweringAction Recognition

Using an external website with a specialised AI Trainer Account. This project requested a long conversation with an LLM under development (10+ turns) with at least 20+ constraints per prompt. The goal of this project was to get the LLM to fail instruction following in long conversations and with a large number of parameters. The LLM produced 2 responses per prompt, and I had to rate it along different rating axes, such as factuality, writing quality, formatting, appropriate length and which followed the instructions better. This project takes 5-7+ hours to complete fully.

Using an external website with a specialised AI Trainer Account. This project requested a long conversation with an LLM under development (10+ turns) with at least 20+ constraints per prompt. The goal of this project was to get the LLM to fail instruction following in long conversations and with a large number of parameters. The LLM produced 2 responses per prompt, and I had to rate it along different rating axes, such as factuality, writing quality, formatting, appropriate length and which followed the instructions better. This project takes 5-7+ hours to complete fully.

2025

Teaching AI To Be More Honest - Document Citation

Internal Proprietary ToolingDocumentQuestion AnsweringObject Detection

This project (hosted on an external hosting site using a specialised AI Training account) was about training a AI citation bot in development at the time. The task was to upload 5-10+ pdf documents to the AI and give it difficult prompts that required accurate citations from multiple documents to arrive at a satisfactory answer. This AI was in development for Higher Education and Research purposes, which most of the documents used for this training were aligned with - although we were free to give it documents from all sorts of topics (within guidelines, of course).

This project (hosted on an external hosting site using a specialised AI Training account) was about training a AI citation bot in development at the time. The task was to upload 5-10+ pdf documents to the AI and give it difficult prompts that required accurate citations from multiple documents to arrive at a satisfactory answer. This AI was in development for Higher Education and Research purposes, which most of the documents used for this training were aligned with - although we were free to give it documents from all sorts of topics (within guidelines, of course).

2025 - 2025

Factuality Evaluation

Data Annotation TechTextClassificationEvaluation Rating

This factuality project was long-running on the DataAnnotationTech platform. It involved analysing a chatbot's long response to a prompt that was seeking factual information. The response was rated one of these labels after analysis: Inaccurate, Unsupported, Disputed or Accurate. All factual claims had to be checked through external online search, and external URLs were needed to confirm final ratings - as well as 5+ sentences for explanation of the rater's result.

This factuality project was long-running on the DataAnnotationTech platform. It involved analysing a chatbot's long response to a prompt that was seeking factual information. The response was rated one of these labels after analysis: Inaccurate, Unsupported, Disputed or Accurate. All factual claims had to be checked through external online search, and external URLs were needed to confirm final ratings - as well as 5+ sentences for explanation of the rater's result.

2023 - 2025

Domain Expertise - Hiberno-English

Data Annotation TechTextClassificationQuestion Answering

This project aimed to train and correct mistakes of an under-development LLM training in the Hiberno-English dialect. My job was to have multi-turn conversations with the LLM in certain categories (e.g. creative writing, roleplaying, fact-seeking, classification etc), and reward good use of Hiberno-English while correcting bad use of Hiberno-English.

This project aimed to train and correct mistakes of an under-development LLM training in the Hiberno-English dialect. My job was to have multi-turn conversations with the LLM in certain categories (e.g. creative writing, roleplaying, fact-seeking, classification etc), and reward good use of Hiberno-English while correcting bad use of Hiberno-English.

2024 - 2024

Education

Work History