Model Card for Model ID
GPT-Neo 125M Customer Support Chatbot
This is a lightweight, fine-tuned large language model designed to simulate a polite, helpful customer support agent. It was trained using Low-Rank Adaptation (LoRA) on the Bitext Customer Support dataset.
Model Details
Model Description
This model is a fine-tuned version of EleutherAI/gpt-neo-125M. It has been specifically trained to handle common customer support queries such as account recovery, order tracking, cancellations, and basic troubleshooting. The model excels at adopting an empathetic, professional tone and structuring responses with clear, step-by-step instructions. pushed on the Hub. This model card has been automatically generated.
- Developed by: [Nourhan Amro/nourhan214]
- Model type: [Causal Language Model (Transformer)]
- Language(s) (NLP): [English]
- License: [MIT]
- Finetuned from model [optional]: [EleutherAI/gpt-neo-125M]
Model Sources [optional]
- Repository: [More Information Needed]
- Paper [optional]: [More Information Needed]
- Demo [optional]: [More Information Needed]
Uses
Direct Use
This model is intended for experimental and educational use to demonstrate how small-parameter models (125M) can learn domain-specific tone and formatting (like customer support personas and template placeholders) through PEFT/LoRA fine-tuning.
Out-of-Scope Use
Due to its small parameter size, this model should not be used in a live production environment without human supervision. The model is prone to hallucinations and may generate factually incorrect troubleshooting steps or get stuck in repetitive empathetic loops. It is not capable of complex reasoning or accessing real-time database information.
Bias, Risks, and Limitations
- Hallucinations: The model relies heavily on pattern matching. It may invent non-existent UI buttons, web pages, or company policies when attempting to solve a user's problem.
- Repetition: At lower temperatures, the model may repeat empathetic filler phrases (e.g., "I'm here to help you every step of the way") multiple times in a single response.
- Context Loss: For multi-turn conversations, the model may quickly lose track of the original user intent.
Recommendations
If using this model for testing, we highly recommend setting a low generation temperature (e.g., 0.3 to 0.4) to reduce severe hallucinations, while using a max_new_tokens of 256 to prevent mid-sentence cutoffs.
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
How to Get Started with the Model
Use the code below to get started with the model.
[python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "nourhan214/gptneo125m-bitext-customer-support"
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)
prompt = "I placed an order yesterday, but I haven't received a confirmation email."
input_text = f"User: {prompt}\nAssistant:"
inputs = tokenizer(input_text, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=256,
pad_token_id=tokenizer.eos_token_id,
temperature=0.3,
do_sample=True,
top_p=0.9
)
full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
assistant_reply = full_response.split("Assistant:")[-1].strip()
print(assistant_reply)]
Training Details
Training Data
[The model was trained on a downsampled subset (15,000 training rows, 2,000 validation rows) of the Bitext Customer Support LLM Chatbot Training Dataset.]
Training Procedure
Preprocessing []
[The dataset was formatted into a conversational prompt structure:
User: {instruction}\nAssistant: {response}<|endoftext|>
The sequence length was capped at 256 tokens.]
Training Hyperparameters
- Training regime: [
- The model was trained using LoRA with the following configurations:
-LoRA Rank (r): 16
-LoRA Alpha: 32
-Target Modules: q_proj, k_proj, v_proj, out_proj
-Dropout: 0.05
]
- Training Arguments: [
- Training regime: fp16 mixed precision
-Epochs: 3
-Learning Rate: 2e-4
-Per-device batch size: 4
-Gradient accumulation steps: 4
-Weight decay: 0.01
-Warmup steps: 100
]
Speeds, Sizes, Times [optional]
[Hardware: NVIDIA Tesla T4 GPU
Training Time: ~31 minutes
Final Training Loss: 1.28
Final Validation Loss: 1.29]
