Charles Y.

Key Skills

Software

Data Annotation Tech

Mercor

Mindrift

OneForma

Remotasks

Scale AI

SuperAnnotate

Snorkel AI

Top Subject Matter

Artificial Intelligence & Machine Learning (LLM Benchmarking / RLHF)

Advanced Mathematics (Calculus, Physics Dynamics, & Logic)

E-commerce (Digital Dropshipping & Product Showcase Design)

Top Data Types

Computer Code Programming

Text

Video

Top Task Types

RLHF

Transcription

Computer Programming/Coding

Prompt + Response Writing (SFT)

Question Answering

Text Generation

Freelancer Overview

I have extensive experience in high-level AI training and quality evaluation, specifically focusing on Reinforcement Learning from Human Feedback (RLHF) and the development of robust benchmarking frameworks for Large Language Models (LLMs). My work involves a deep technical understanding of audio script tagging, language evaluation, and the creation of detailed rubrics for assessing AI output. I have contributed to complex projects like "Achilles" and "Styx," where I specialized in grounding model responses and evaluating multifaceted linguistic tasks. Beyond operational tasks, my academic research at Peking University regarding computational logic, combined with my published work in EMNLP and CVPR, allows me to approach data labeling with a sophisticated understanding of how training data architecture directly impacts model performance and scalability.

Labeling Experience

Deep Learning Data Annotation for CNN Architectures

TextRLHF

This project involved the high-granularity annotation of visual and linguistic datasets to support the development of domain-specific Convolutional Neural Network (CNN) architectures and cross-modal classification models. The scope encompassed the labeling of thousands of images for object detection and the curation of sentiment-based text datasets for conversion into image-matrix representations, as detailed in my EMNLP and CVPR publications. My specific tasks included defining bounding box parameters for real-time IOT object recognition and establishing a systematic mapping between natural language sentiment and visual features. To ensure rigorous quality, I adhered to iterative cross-validation measures and utilized multi-stage verification to maintain high inter-annotator agreement and training data integrity. This foundational work contributed to the successful filing of two U.S. patents and the development of ultra-power-efficient accelerators achieving 9.3 TOPS/Watt.

2018 - 2019

Education

Work History