Zeming Z.

Key Skills

Software

Don't disclose

Other

Top Subject Matter

AI model output evaluation and preference labeling (LLM training)

Generative video evaluation and rubric-based rating

Generative audio quality evaluation and emotion-tag based labeling

Top Data Types

Text

Video

Audio

Image

Top Task Types

RLHF

Emotion Recognition

Freelancer Overview

AI Training Contributor — Freelance / Independent (LLM preference evaluation and RLHF-style labeling). Brings 3+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Don't disclose and Other. Education includes Bachelor of Fine Arts, Tianjin Polytechnic University (2022). AI-training focus includes data types such as Text, Video, and Audio and labeling workflows including RLHF, Evaluation, and Rating.

Labeling Experience

AI Training Contributor — Freelance / Independent (Audio rubric design and emotion-aware evaluation)

OtherAudioEmotion Recognition

Created evaluation criteria for voice synthesis outputs focusing on emotion-tag execution and tone control. Assessed whether emotion-tag dosage was reflected in delivery and monitored for advertising-tone leakage. Used rubric-driven checks to validate voice quality for English voice generation workflows. • Evaluated ElevenLabs v3 results against emotion-tag expectations. • Detected undesired tonal artifacts such as ad-tone leakage. • Verified correct timing and delivery alignment with tags like pause. • Reused labeling logic across repeated generation scenarios.

2024 - Present

AI Training Contributor — Freelance / Independent (Video rubric design and quality evaluation)

OtherVideo

Designed reusable evaluation rubrics for generated video outputs to score and validate adherence to motion and framing constraints. Defined testable success criteria such as camera-lock behavior, single-action requirements, and a short temporal window for optimal results. Used these rubrics to assess generations from agentic video pipelines and capture edge cases as reproducible failures. • Evaluated Seedance 2.0 video outputs using structured rubrics. • Assessed camera-lock and continuity properties. • Checked action complexity for single-action execution. • Tuned rubric expectations for a 5-second target segment.

2024 - Present

AI Training Contributor — Freelance / Independent (LLM preference evaluation and RLHF-style labeling)

Don't discloseTextRLHF

Provided structured preference feedback and written rationales to evaluate model outputs and guide AI training using RLHF-style judgments. Performed comparative ranking across generated responses and tracked quality issues such as hallucinations and refusal quality. Used multimodal evaluation practices to ensure consistency across text-driven agentic and generative workflows. • Compared multiple outputs from Claude (Opus/Sonnet) and OpenAI Codex. • Produced prompt iteration notes and rationale-driven feedback. • Reviewed refusals and hallucination likelihood based on rubric criteria. • Documented recurring failure modes as test cases.

2024 - Present

Education

Work History