sarcastic-bakery-lora

Sarcastic Bakery Chatbot — LoRA Adapter

This repository contains LoRA adapter weights for a fine-tuned large language model that produces sarcastic yet polite responses, while strictly adhering to a bakery-only assistant role.

The model was fine-tuned as part of an academic project to demonstrate parameter-efficient fine-tuning (PEFT) of LLMs.

---

Model Details

Model Description

• Developed by: Harshika Dewani
• Model type: LoRA adapter for causal language model
• Language(s): English
• Base model: unsloth/llama-3-8b-bnb-4bit
• Fine-tuning method: LoRA (Low-Rank Adaptation)
• Quantization: 4-bit (QLoRA)
• License: Same as base model (LLaMA 3 license)

This model modifies the conversational style and tone of the base model to introduce sarcasm while preserving politeness and role adherence.

---

Intended Uses

Direct Use

This model is intended to be loaded as a LoRA adapter on top of the base LLaMA-3 model to generate sarcastic bakery-themed responses.

Out-of-Scope Use

• Medical, legal, financial, or technical advice
• Non-bakery domain conversations
• Malicious or harmful content generation

---

Training Details

Training Data

• Custom instruction–response dataset
• Domain: Bakery customer support
• Focus:
  • Sarcastic tone
  • Polite refusals
  • Strict role adherence
• Dataset size: ~100–300 curated examples
• Format: Instruction → Response pairs

Training Procedure

• Framework: Unsloth + TRL
• Trainer: SFTTrainer
• Fine-tuning strategy: Supervised Fine-Tuning (SFT) with LoRA
• Only LoRA adapter weights were trained; base model weights were frozen.

Training Hyperparameters

• Batch size (per device): 2
• Gradient accumulation steps: 4
• Learning rate: 2e-4
• Max training steps: 100
• Precision: FP16 / BF16 (hardware dependent)
• GPU: Free Google Colab GPU

---

Evaluation

Evaluation Approach

The model was evaluated qualitatively using:

• Before vs after comparison with the base model
• Manual inspection of:
  • Sarcasm presence
  • Politeness
  • Role adherence

Example behavior change:

Prompt:

Do you sell pizza?

• Base model: Neutral refusal
• Fine-tuned model: Sarcastic, bakery-focused refusal

---

How to Use the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = "unsloth/llama-3-8b-bnb-4bit"
adapter = "your-username/sarcastic-bakery-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    load_in_4bit=True,
    device_map="auto"
)

model.load_adapter(adapter)

prompt = "Do you sell pizza?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))