persona_maker

Persona Generator on Qwen3 14B

Model summary

<persona_maker> is a fine-tuned version of Qwen3 14B (reasoning) that turns a short job title, role, or occupation into a full, ChatGPT-style persona prompt.

• Base model: Qwen3 14B (reasoning variant)
• Fine-tuning framework: Unsloth
• Objective: Given an input like "Ethical Hacker" or "Senior Data Scientist in fintech", generate a detailed “I want you to act as…” style persona prompt that can be used directly as a system / role prompt for chat models.
• Primary language: English

What the model does

Input (example):

Ethical Hacker

Output (example):

I want you to act as an Ethical Hacker. I will provide you with details of a system or network that I want you to test for vulnerabilities. Your task is to use your hacking skills to find any weaknesses in the system and report them to me in a clear and concise manner. My first request is "I need you to test my company's website for any security flaws."

The model generalizes this pattern to any occupation or role: “Family Doctor”, “Product Manager at a SaaS startup”, “Kubernetes SRE”, “High school math teacher”, etc.

---

Intended use

Intended use cases

• Prompt engineering helpers
  • Quickly generate structured persona prompts for downstream LLMs.
  • Create consistent “role” templates for agents / tools.
• App / product integrations
  • “Prompt template builder” features in chat apps.
  • Internal utilities for generating standardized system prompts for different teams (support, sales, research, etc.).
• Educational / experimentation
  • Explore how different personas are framed and what instructions a model might follow.

Input:
A short description of a role, job title, or persona.
Examples:

• "Ethical Hacker"
• "Quantitative Researcher in an Indian equity hedge fund"
• "Career coach for software engineers"

Output:
A single, long-form persona prompt in natural language, typically starting with “I want you to act as …” plus context and first request.

Out-of-scope / misuse

This model does not:

• Perform actual reasoning or execution as the persona (it only generates the persona prompt).
• Guarantee safety or legality of the persona description, especially for sensitive roles like:
  • Hacking / security testing
  • Medical, legal, or financial advisors
  • Anything that might encourage harmful, illegal, or unethical behavior

You are responsible for:

• Adding safety filters or post-processing.
• Enforcing usage policies in your application (e.g. blocking obviously harmful persona requests).
• Complying with the underlying Qwen license and your own deployment environment’s policies.

---

Training details

Base model

• Architecture: Qwen3 14B (reasoning variant)
• Type: Chat / instruction-following LLM

Fine-tuning

• Framework: Unsloth
• Method: Parameter-efficient fine-tuning (e.g. LoRA/QLoRA-style) on persona-style data derived from:
  • Dataset: fka/awesome-chatgpt-prompts

The dataset consists of prompts like:

I want you to act as an Ethical Hacker. I will provide you with details of a system or network that I want you to test for vulnerabilities...

These were adapted into a (role description → persona prompt) style training task so that the model learns to map a short role name to a full persona description.

⚠️ Exact hyperparameters (epochs, learning rate, LoRA ranks, etc.)
Fill this in with your real values if you want:

• epochs: <n>
• learning_rate: <lr>
• max_seq_length: <seq_len>
• optimizer: <adamw / etc.>

---

How to use

Basic usage (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "agrashu/persona_maker"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
)

def generate_persona(role: str, max_new_tokens: int = 256):
    prompt = f"Role: {role}\nPersona Prompt:"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=0.9,
        temperature=0.7,
    )
    text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return text

print(generate_persona("Ethical Hacker"))