Ephraimmm/nicer_customer_service_model

Nicer Customer Service Model

Overview

This model is a LoRA fine-tune of unsloth/meta-llama-3.1-8b-unsloth-bnb-4bit (a 4-bit quantized Llama 3.1 8B base) adapted for customer-service style dialogue. It was trained to better recognize common customer intents and to respond in a more helpful, courteous ("nicer") tone suitable for support chatbot use cases. Only the LoRA adapter weights are hosted in this repository; they are applied on top of the base model at inference time.

Training Details

Detail	Value
Base model	unsloth/meta-llama-3.1-8b-unsloth-bnb-4bit (Llama 3.1 8B, 4-bit bitsandbytes quantization)
Fine-tuning method	LoRA (PEFT), adapter weights only (adapter_model.safetensors)
LoRA rank (r)	16
LoRA alpha	16
LoRA dropout	0
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Bias	none
Task type	CAUSAL_LM
Training frameworks	Unsloth + Hugging Face TRL + PEFT, reported by the author to run ~2x faster via Unsloth
License	Apache 2.0
Language	English

Exact training step/epoch counts, dataset details, and hardware are not recorded in this repository (no trainer_state.json or training logs were published), so they are not claimed here.

Intended Use

• Drafting or generating responses for customer-support chat interfaces
• Recognizing and responding to common customer intents (e.g., order status, complaints, general inquiries)
• As a starting point/adapter for further fine-tuning on a specific company's support domain

This model is intended as a lightweight, task-oriented assistant for customer-service style conversations, not as a general-purpose chat assistant.

How to Use

Because this repository contains a PEFT LoRA adapter (not a merged model), load the base model first and then attach the adapter:

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

base_model_id = "unsloth/meta-llama-3.1-8b-unsloth-bnb-4bit"
adapter_id = "Ephraimmm/nicer_customer_service_model"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

prompt = "Customer: My order hasn't arrived yet, what should I do?\nAgent:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(output[0], skip_special_tokens=True))

For faster loading/inference with 4-bit quantization, you can also load the base model with Unsloth's FastLanguageModel and attach this adapter on top.

Limitations

• This is a narrow, domain-focused fine-tune (customer-service dialogue) and has not been evaluated against standard NLP or chat benchmarks — no benchmark scores are published for this model.
• No evaluation metrics, dataset documentation, or training logs are included in this repository; performance has not been independently quantified.
• As with any LLM fine-tune, the model can still produce inaccurate, incomplete, or inappropriate responses and should be reviewed by a human before use in production customer-facing systems.
• Inherits the general limitations and biases of the underlying Llama 3.1 8B base model.

Author

Developed by Ephraimmm.