Nanami Antoku

Key Skills

Software

Toloka

Labelbox

Mindrift

Mercor

Top Subject Matter

Natural Language Processing

Legal and Public Documents

Social Media

Top Data Types

Document

Audio

Text

Image

Top Task Types

Classification

Evaluation/Rating

Audio Recording

Transcription

Entity (NER) Classification

Question Answering

Text Generation

RLHF

Prompt + Response Writing (SFT)

Freelancer Overview

Data Annotator. Core strengths include Internal and Proprietary Tooling. AI-training focus includes data types such as Text and labeling workflows including Classification.

Labeling Experience

Audio Annotation (Transcription / QA)

LabelboxAudioTranscription

Worked on transcription and quality assurance tasks for Japanese audio datasets through projects with OpenTrainAI. My primary responsibility was to accurately transcribe provided audio data into text based on client-specific guidelines. In addition, I performed QA tasks to evaluate whether the transcriptions met the required standards and specifications. Key aspects of my work included: Accurately capturing non-verbal elements such as fillers, pauses, and hesitations Reflecting language-specific nuances and colloquial expressions in Japanese Applying consistent labeling decisions based on predefined guidelines Through this work, I ensured high-quality and standardized data annotation with strong adherence to reproducible rules and criteria.

Worked on transcription and quality assurance tasks for Japanese audio datasets through projects with OpenTrainAI. My primary responsibility was to accurately transcribe provided audio data into text based on client-specific guidelines. In addition, I performed QA tasks to evaluate whether the transcriptions met the required standards and specifications. Key aspects of my work included: Accurately capturing non-verbal elements such as fillers, pauses, and hesitations Reflecting language-specific nuances and colloquial expressions in Japanese Applying consistent labeling decisions based on predefined guidelines Through this work, I ensured high-quality and standardized data annotation with strong adherence to reproducible rules and criteria.

2025 - 2025

Japanese STEM text annotator

TolokaDocumentEvaluation Rating

Worked on evaluating text generation datasets in the STEM domain through projects with Outlier. My responsibilities included assessing the quality of AI-generated text based on provided datasets and assigning evaluation ratings accordingly. Evaluations were conducted with consideration of domain-specific knowledge, focusing on the following criteria: Accuracy: Whether the content was factually correct and aligned with given conditions Coherence / Reasoning: Logical consistency and clarity of explanations Overall Quality: Readability and appropriateness of the generated text Through this work, I contributed to improving the overall quality of training data by providing structured and consistent evaluation.

Worked on evaluating text generation datasets in the STEM domain through projects with Outlier. My responsibilities included assessing the quality of AI-generated text based on provided datasets and assigning evaluation ratings accordingly. Evaluations were conducted with consideration of domain-specific knowledge, focusing on the following criteria: Accuracy: Whether the content was factually correct and aligned with given conditions Coherence / Reasoning: Logical consistency and clarity of explanations Overall Quality: Readability and appropriateness of the generated text Through this work, I contributed to improving the overall quality of training data by providing structured and consistent evaluation.

2024 - 2024

Data annotation (Outlier)

MindriftTextRLHFClassification

Role: AI Annotation (RLHF Evaluation) Worked on evaluating text outputs related to AI-generated prompts and corresponding images through projects obtained via Outlier. Specifically, I assessed multiple AI-generated image descriptions based on images created from user-provided Japanese input text. My role involved comparing these descriptions and selecting the most appropriate one. Evaluations were conducted based on the following criteria: Accuracy: Alignment between the image and its description Safety: Presence or absence of harmful or inappropriate content Instruction Following: Degree to which the description adhered to the intent of the input text Through this process, I performed quality assessment and ranking of AI outputs, contributing to the improvement of model performance.

Role: AI Annotation (RLHF Evaluation) Worked on evaluating text outputs related to AI-generated prompts and corresponding images through projects obtained via Outlier. Specifically, I assessed multiple AI-generated image descriptions based on images created from user-provided Japanese input text. My role involved comparing these descriptions and selecting the most appropriate one. Evaluations were conducted based on the following criteria: Accuracy: Alignment between the image and its description Safety: Presence or absence of harmful or inappropriate content Instruction Following: Degree to which the description adhered to the intent of the input text Through this process, I performed quality assessment and ranking of AI outputs, contributing to the improvement of model performance.

2025 - Present

RLHF Data Creation & Evaluation Design

DocumentRLHF

This work was conducted through a project with Mercor. I was responsible for designing and creating domain-specific datasets (in my case, chemistry) and developing prompts that required the AI to perform analytical tasks based on those datasets. For evaluation, I predefined multiple criteria and instructed the model accordingly. I then assessed whether the generated responses adhered to these criteria. The evaluation framework itself was designed from multiple perspectives, incorporating domain expertise. I iteratively refined and tightened the criteria to the point where only highly accurate and well-reasoned responses could fully satisfy them. This was a highly demanding task, with approximately 30 to 60 minutes allocated per response to ensure thorough analysis and evaluation.

This work was conducted through a project with Mercor. I was responsible for designing and creating domain-specific datasets (in my case, chemistry) and developing prompts that required the AI to perform analytical tasks based on those datasets. For evaluation, I predefined multiple criteria and instructed the model accordingly. I then assessed whether the generated responses adhered to these criteria. The evaluation framework itself was designed from multiple perspectives, incorporating domain expertise. I iteratively refined and tightened the criteria to the point where only highly accurate and well-reasoned responses could fully satisfy them. This was a highly demanding task, with approximately 30 to 60 minutes allocated per response to ensure thorough analysis and evaluation.

2026 - 2026

Education

Work History