Hongli W.

Key Skills

Software

Label Studio

Other

Argilla

Top Subject Matter

LLM alignment

RLHF ranking

text safety evaluation

Top Data Types

Text

Top Task Types

RLHF

Red Teaming

Prompt + Response Writing (SFT)

Classification

Freelancer Overview

AI Trainer – Self-Trained & Calibrated Projects (RLHF preference ranking and model response scoring). Core strengths include Label Studio, Other, and Microsoft Excel. Education includes Junior College Diploma, N/A (2018). AI-training focus includes data types such as Text and labeling workflows including RLHF, Red Teaming, and Evaluation.

Labeling Experience

AI Trainer – Self-Trained & Calibrated Projects (Content safety labeling and risk classification)

ArgillaTextClassification

Performed full-process content safety labeling by classifying text risks such as violence, misleading information, and privacy leakage. Applied labeling work in line with international AI data security specifications. Ensured safe dataset construction through consistent categorization of risk-related content. • Labeled and classified high-risk safety categories • Identified violence, misleading content, and privacy leakage risks • Followed international AI data security labeling specifications • Produced consistently categorized safety annotations

2025 - Present

AI Trainer – Self-Trained & Calibrated Projects (Chinese SFT prompt-response curation)

TextPrompt + Response Writing (SFT)

Created high-quality Chinese SFT dialogue pairs by selecting and screening non-repetitive, non-vague, high-value text examples for general large models. Converted and unified the dataset into a standard JSON format and enriched multi-scenario dialogue samples. Curated prompt-response data with attention to consistency and dataset usability for fine-tuning. • Authored Chinese SFT prompt-response dialogue pairs • Screened out repetitive, vague, and low-value texts • Normalized outputs into standard JSON • Expanded dialogues across multiple scenarios for SFT

2025 - Present

AI Trainer – Self-Trained & Calibrated Projects (Math reasoning answer validation)

Text

Validated step-by-step chain-of-thought math answers for middle and high school subjects by pinpointing miscalculations, logical flaws, and incomplete derivations. Rewrote standardized correct reasoning to improve the training data quality and correctness. Verified mathematical reasoning consistency as part of dataset calibration. • Checked CoT correctness for math solutions • Identified calculation errors and logical gaps • Rewrote correct standardized reasoning content • Ensured derivation completeness for training data quality

2025 - Present

AI Trainer – Self-Trained & Calibrated Projects (LLM red-teaming and security risk recording)

OtherTextRed Teaming

Conducted LLM red-team simulation work using targeted adversarial prompt design to detect model security loopholes. Sorted and recorded risky output cases into standardized forms for downstream evaluation and mitigation. Focused on identifying and documenting model security and safety failures during adversarial testing. • Designed adversarial prompts to probe security loopholes • Performed red-team simulations of LLM behavior • Sorted and recorded risky outputs in standardized templates • Used findings to support model safety evaluation workflows

2025 - Present

AI Trainer – Self-Trained & Calibrated Projects (RLHF preference ranking and model response scoring)

Label StudioTextRLHF

Performed RLHF preference ranking by scoring LLM responses across helpfulness, factual accuracy, logical coherence, and safety dimensions, and labeling issues like hallucinations, bias, and invalid replies. Supported reward model iteration by recording risky outputs in a consistent manner aligned to official rubrics. Maintained stable inter-annotator agreement and uniform labeling standards throughout the project cycle. • Helpfulness, factual accuracy, logical coherence, and safety scoring for model responses • Labeling hallucinations, biased outputs, and invalid replies • Standardized documentation of risky or non-compliant cases • Adherence to official scoring rubrics with quality consistency

2025 - Present

Education

Work History