Harsha-AI-Qwen3

Model Card for Harsha-AI-Qwen3

Model Details

Model Description

This is a custom AI persona model fine-tuned to simulate the conversational style, professional background, and design philosophy of Harsha Karunarathna. It was developed using a Parameter-Efficient Fine-Tuning (PEFT) approach with LoRA on the lightweight Qwen3-0.6B base model. The model is optimized for 4-bit precision, making it highly efficient for deployment and inference.

• Developed by: Harsha Karunarathna
• Model type: Causal Language Model (Fine-tuned via LoRA)
• Language(s) (NLP): English
• Finetuned from model: Qwen/Qwen3-0.6B

Model Sources

• Repository: https://huggingface.co/Harsha25ai/Harsha-AI-Qwen3

Uses

Direct Use

This model is intended to be used as a digital conversational persona. Users can interact with the model to ask questions about Harsha's design philosophy (e.g., regarding Pixera or Harsha Store), software engineering concepts, and project experiences.

Out-of-Scope Use

This model should not be used as a source of absolute factual truth outside of its trained persona context. It is not designed for generating medical, legal, or financial advice. Furthermore, it should not be used for malicious purposes, harassment, or generating toxic content.

Bias, Risks, and Limitations

• Hallucinations: Like all LLMs, this model may occasionally generate plausible-sounding but factually incorrect information (hallucinations). It may invent experiences or opinions that the real Harsha Karunarathna does not hold.
• Data Privacy & Prompt Injection: Because the model was trained on a custom conversational dataset, advanced users might use prompt injection techniques to extract specific phrasing or underlying data from the training set.
• Persona Drift: Depending on the prompt, the model may sometimes drop its designated persona and answer as a generic AI assistant.

Recommendations

Users should be aware that this is an experimental AI representation and its responses are automated. Do not treat the model's outputs as legally binding statements or exact representations of the creator's personal life.

How to Get Started with the Model

Use the code below to load and interact with the model:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

model_id = "Qwen/Qwen3-0.6B"
adapter_id = "Harsha25ai/Harsha-AI-Qwen3"

# Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

# Load Base Model
base_model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

# Attach LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)

# Chat Function
messages = [
    {"role": "system", "content": "You are the AI persona of Harsha Karunarathna. Speak as Harsha."},
    {"role": "user", "content": "What is your design philosophy for Pixera?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))