itsrenzosamaaa/robby-chat-model

Robby Chat Model 🤖

A personal fine-tuned language model exploring conversational personality, humor, and natural chat.

Robby is an experimental fine-tune of Qwen2.5-3B-Instruct, created to explore how a small language model can learn an individual's communication style, humor, and casual conversational patterns.

This project is part of my personal AI engineering journey, combining dataset preparation, model fine-tuning, and experimentation with open-weight language models.

✨ Project Goals

• Personality imitation: Explore how a model can reproduce distinctive conversational patterns.
• Casual conversation: Experiment with natural, informal responses rather than overly formal assistant-style answers.
• Humor and slang: Explore humor, playful banter, and everyday conversational language.
• Efficient fine-tuning: Adapt an existing language model rather than training one from scratch.
• Personal AI experimentation: Learn more about dataset quality, fine-tuning, inference, and model evaluation.

🧠 Model Details

Property	Details
Model name	Robby Chat Model
Base model	Qwen2.5-3B-Instruct
Parameters	Approximately 3 billion
Fine-tuning framework	Unsloth and Hugging Face TRL
Task	Conversational text generation
Model format	Safetensors
Intended use	Personal experimentation and casual conversation

🛠️ Training Approach

The model was fine-tuned using conversational data prepared for this personal project.

The workflow involved:

1. Preparing and processing conversational examples.
2. Cleaning and organizing the dataset for training.
3. Fine-tuning an existing instruction-tuned model with the Unsloth and Hugging Face ecosystem.
4. Saving the resulting model checkpoint for experimentation and inference.

The focus is on teaching conversational patterns through examples rather than relying entirely on a long system prompt.

💬 Example Use Cases

Robby is intended for experimentation with:

• Casual chat and playful banter.
• Informal English and Taglish conversations.
• Humor, slang, and conversational expression.
• Personal AI assistants and Discord bot prototypes.
• Studying how fine-tuning affects a model's response style.

These are intended use cases, not guarantees of performance.

🚀 Getting Started

You can load the model with the Hugging Face Transformers library, provided the repository contains a complete Transformers-compatible checkpoint.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "itsrenzosamaaa/robby-chat-model"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Bro, kumusta ka na?"
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,
        do_sample=True,
        temperature=0.8,
        top_p=0.9,
    )

response = outputs[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

Note: Hardware requirements depend on the model's precision and available memory. Adjust the loading configuration if your environment cannot accommodate the checkpoint.

⚠️ Limitations

• This is an experimental personal fine-tune, not a general-purpose production assistant.
• Responses may be repetitive, inconsistent, inaccurate, or contextually inappropriate.
• Conversational style imitation does not guarantee factual accuracy or strong reasoning.
• Performance on standard benchmarks may not reflect how convincingly the model reproduces a particular personality.
• The model has not been independently benchmarked as part of this project unless results are published separately.

🔒 Dataset and Privacy

The underlying conversational dataset is not included in this repository.

This project is intended to experiment with conversational style while respecting the privacy of people whose messages may appear in source conversations. Do not assume that model outputs are guaranteed to exclude information memorized during training.

📜 License

The repository currently lists the Apache-2.0 license. Review the base model's license and the rights associated with your training data before redistributing model weights or using them commercially.

🙌 Acknowledgments

• Qwen for the base model.
• Unsloth for efficient fine-tuning tools.
• Hugging Face for the model-hosting and machine-learning ecosystem.

---

Built as a personal AI engineering experiment — one conversation at a time.