gpt-nq-prompt-generator

gpt-nq-prompt-generator

This is a fine-tuned version of the GPT-2 medium variant that was developed to enhance prompt engineering capabilities. Its primary capability is to generate contextualized and detailed prompts based on job titles using transfer knowledge. Given a job title as input, such as "Software Engineer" or "Data Scientist", the model outputs a well-crafted description that highlights the roles, responsibilities, and typical tasks associated with that job.

Output Examples

Intended Use

The sgarbi/gpt-nq-prompt-generator is designed with specificity in mind.

The model's main purpose is to generate detailed AI prompts for an array of professional roles, providing users with in-depth insights into the responsibilities and nuances of different job roles.

Potential Applications:

1. Educational Platforms: Serve as a guiding tool for platforms offering insights into various job roles.

2. Job Portals and Recruitment: Assist recruiters in crafting comprehensive job descriptions.

3. Chatbots and Virtual Assistants: Enhance chatbot systems by offering users detailed information about various professions.

Licensing: This model is released under the MIT license, in alignment with GPT-2's licensing provisions. During its fine-tuning, the Natural Questions (NQ) dataset, last known to be under a Creative Commons Attribution 4.0 International License as of January 2022, was utilized. Users are encouraged to keep abreast of the latest licensing terms associated with the datasets and tools they engage with.

How To Use

1. Input Format: Always input the desired role or job title as a straightforward prompt. For example, "Software Engineer" or "Nurse Practitioner".
2. Tag Use: While the model has been trained with an array of job titles, it recognizes them best when they are input without additional context or embellishments.
3. Result: The model will provide a synthesized description, drawing from its training, to offer detailed information about the specified role.

Note:

While the model recognizes a diverse range of job titles, it's always possible that some niche or highly specialized roles might receive less detailed or generic outputs. In such cases, it might be helpful to slightly modify the input or provide a broader category of the job title.

from transformers import GPT2Tokenizer, GPT2LMHeadModel
import torch


tokenizer = GPT2Tokenizer.from_pretrained('sgarbi/prompt_generator')
tokenizer.pad_token = tokenizer.eos_token

model = GPT2LMHeadModel.from_pretrained('sgarbi/prompt_generator')

def query_model(input_text):
    """Query the model and get a generated response."""
    formatted_input = f"<|title|>{input_text}<|prompt|>"
    input_ids = tokenizer.encode(formatted_input, return_tensors="pt")
    attention_mask = torch.ones(input_ids.shape, dtype=torch.long)

    # Generate a response
    output = model.generate(input_ids, attention_mask=attention_mask, max_length=256, 
                            pad_token_id=tokenizer.eos_token_id, temperature=0, top_k=50, 
                            top_p=0.95, num_return_sequences=1)

    # Decode the output
    decoded = tokenizer.decode(output[0], skip_special_tokens=True)

    tag_index = decoded.find('<|prompt|>')
    return decoded[tag_index + len('<|prompt|>'):]

# Example use
print(query_model("Space X Engineer"))

Using in Transformers:

from transformers import pipeline

pipe = pipeline("text-generation", model="sgarbi/gpt-nq-prompt-generator")
pipe('<|title|>Sales Manager<|prompt|>')

Limitations

• The model's responses are rooted in its training data. While it has knowledge of a wide range of professional roles, there might be some roles it is less familiar with.
• The descriptions are synthetically generated. For critical applications, users should validate the content.

Training Data

The model was fine-tuned on a combination of the NQ (Natural Questions) dataset and a proprietary dataset. The NQ dataset (https://ai.google.com/research/NaturalQuestions/) was instrumental in teaching the model how to answer questions effectively and enabled several passes for coherent knowledge transfer. The proprietary dataset was synthesized using several advanced prompt engineering techniques with the Microsoft Semantic Kernel (https://learn.microsoft.com/en-us/semantic-kernel/overview/) and GPT-3.5-turbo, ensuring the generation of profession-specific AI prompts.

Evaluation

The model's training progress was monitored using a loss metric. The plot showcasing the trend of the training loss over the steps can be inserted here. The loss decreases initially and then stabilizes, indicating that the model is learning and converging.

Compute Infrastructure

Google Collaboratory (https://colab.research.google.com/)

Hardware

A100

Ethics and Bias

Users should be aware that no model is entirely free from biases. We encourage users to interpret its outputs with this in mind and report any issues they encounter.

Licensing

This model is released under the MIT License, aligning with OpenAI's licensing terms.

Github

https://github.com/ESgarbi/gpt-nq-prompt-generator