🤖 Sarcastic-TinyLlama-Adapter

This is a LoRA (Low-Rank Adaptation) adapter for TinyLlama-1.1B-Chat-v1.0.

It was fine-tuned to transform the helpful assistant into a sarcastic, witty, and slightly unhelpful bot. It answers questions correctly but adds a layer of snark.

📝 Model Details

• Base Model: TinyLlama-1.1B-Chat-v1.0
• Fine-Tuning Method: QLoRA (4-bit quantization + LoRA)
• Dataset: Custom "Instruction Tuning" dataset for sarcasm.
• Hardware: Trained on a single T4 GPU (Google Colab).
• Training Time: ~15 minutes.

💻 How to Use

You can load this model using the peft and transformers libraries.

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

# 1. Load Base Model
base_model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
adapter_model_name = "pvs88/Sarcastic-TinyLlama-Adapter"

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# 2. Load Adapter (The Sarcasm)
model = PeftModel.from_pretrained(base_model, adapter_model_name)
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# 3. Chat!
def ask(question):
    prompt = f"<|user|>\n{question}</s>\n<|assistant|>\n"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    outputs = model.generate(
        **inputs, 
        max_new_tokens=50, 
        do_sample=True, 
        temperature=0.7
    )
    print(tokenizer.decode(outputs[0], skip_special_tokens=True))

ask("My internet is broken.")
# Expected Output: "Don't worry, it's always broken. Have you tried staring at the router?"

🛠️ Training Parameters

• LoRA Rank (r): 64
• Epochs: 5
• Learning Rate: 2e-4
• Batch Size: 4

⚠️ Limitations

• This model is for entertainment purposes only.
• It is very small (1.1B), so it may hallucinate facts.
• It is specifically trained to be rude/sarcastic; do not use for customer support!