J. Wicher

Key Skills

Software

Data Annotation Tech

Top Subject Matter

No subject matter listed

Top Data Types

Image

Text

Top Task Types

Classification

Fine Tuning

Prompt Response Writing SFT

Freelancer Overview

I have participated in projects in Data Annotation, mainly assessing the quality of models responses, according to directives and instructions given by employer. I have also participated in projects involving prompt writing and classifying images.

Labeling Experience

Arsenic Project Training - Safety risks

Data Annotation TechTextClassification

This training involved understanding the types of prompts and classifying them according to their level of toxicity (safe, benign, harmful or jailbreak). I also had to recognize the safety risk category (high risk content, sensitive topics, modeling and training risks, illegal behavior, etc). Then, I also had to asses the models response. For example, the model could engage the prompt, shortly decline, decline with a reason or give a disclaimer and then engage. I was also given guidelines as to what the ideal response would be and the characteristics of a harm-free response.

This training involved understanding the types of prompts and classifying them according to their level of toxicity (safe, benign, harmful or jailbreak). I also had to recognize the safety risk category (high risk content, sensitive topics, modeling and training risks, illegal behavior, etc). Then, I also had to asses the models response. For example, the model could engage the prompt, shortly decline, decline with a reason or give a disclaimer and then engage. I was also given guidelines as to what the ideal response would be and the characteristics of a harm-free response.

2024

Project Raven

Data Annotation TechTextQuestion AnsweringText Generation

The project involves fine-tuning the responses of the AI. For that, I had to write a prompt and two answers were given from two different models. From there on, I had to asses the quality of these answers, according to parameters: localization, instruction following and truthfulness. Then, I had to choose which was the best response, how much better was it in relation to the other one and explain issues on both models. After that, I could edit the response that I chose, to make it better when it came to the parameters, if necessary, and explain the changes I made and why. All of this would be done in the context of a conversation with the AI, with a minimum of 2 rounds to a maximum of 6. I was also instruct to make my prompts as natural and real world as possible, to not make too simple, so I would be able to find issues, for example, when there are complex prompts.

The project involves fine-tuning the responses of the AI. For that, I had to write a prompt and two answers were given from two different models. From there on, I had to asses the quality of these answers, according to parameters: localization, instruction following and truthfulness. Then, I had to choose which was the best response, how much better was it in relation to the other one and explain issues on both models. After that, I could edit the response that I chose, to make it better when it came to the parameters, if necessary, and explain the changes I made and why. All of this would be done in the context of a conversation with the AI, with a minimum of 2 rounds to a maximum of 6. I was also instruct to make my prompts as natural and real world as possible, to not make too simple, so I would be able to find issues, for example, when there are complex prompts.

2024

Tango Training

Data Annotation TechTextClassification

I received training to participate in a project involving classifying the models responses to asses whether they complied with the safety policy guidelines or not and why. For example, a model's response might violate these guidelines if it recommends using a specific branded medical product for a health problem or gives legal advice.

I received training to participate in a project involving classifying the models responses to asses whether they complied with the safety policy guidelines or not and why. For example, a model's response might violate these guidelines if it recommends using a specific branded medical product for a health problem or gives legal advice.

2024 - 2024

Prompt writing

Data Annotation TechTextPrompt Response Writing SFT

At the start of my experience in Data Annotation, my tasks involved mainly writing prompts according to certain parameters. Thus, the prompt had to sound natural, give context, have at least three constraints and at least a couple of paragraphs. It had to be a "long prompt". Then, I also had to classify the kind of prompt I had written (asking for advice, brainstorming, open Q&A, closed Q&A, extracting information, etc.). I had also received training in what each of these classifications meant.

At the start of my experience in Data Annotation, my tasks involved mainly writing prompts according to certain parameters. Thus, the prompt had to sound natural, give context, have at least three constraints and at least a couple of paragraphs. It had to be a "long prompt". Then, I also had to classify the kind of prompt I had written (asking for advice, brainstorming, open Q&A, closed Q&A, extracting information, etc.). I had also received training in what each of these classifications meant.

2024 - 2024

Education

Work History