This model is a fine-tuned version of Qwen/Qwen3-1.7B on the Bitext Customer Support dataset. It is designed to act as a helpful customer support agent.

Model Details

Base Model: Qwen/Qwen3-1.7B
Architecture: Qwen3 (1.7B parameters)
Training Method: QLoRA (4-bit quantization with LoRA adapters)
Dataset: Bitext-customer-support-llm-chatbot-training-dataset
Language: English
Code Repository: omid511/qwen3-1.7b-finetune-customer-support

Usage

Loading the Adapter (Default)

python

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

base_model_id = "Qwen/Qwen3-1.7B"
adapter_model_id = "omid5/Qwen3-1.7b-cusomer-support-agent"

# 1. Load Base Model
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# 2. Load Adapters
model = PeftModel.from_pretrained(base_model, adapter_model_id)
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

# 3. Inference
messages = [
    {"role": "user", "content": "I received a defective item, what should I do?"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Loading the Merged Model

The full merged model is available on the merged branch of this repository.

python

model = AutoModelForCausalLM.from_pretrained(
    "omid5/Qwen3-1.7b-cusomer-support-agent",
    revision="merged",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("omid5/Qwen3-1.7b-cusomer-support-agent", revision="merged")

Training Configuration

The model was trained using accelerate and DeepSpeed with the following hyperparameters:

Epochs: 2
Learning Rate: 2e-4 (Cosine Schedule)
Batch Size: 8 (per device train) / 16 (per device eval)
Gradient Accumulation: 2
Max Sequence Length: 400
LoRA Config:
- r: 16
- alpha: 32
- dropout: 0.05
- target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
Quantization: 4-bit (nf4) via bitsandbytes

Training Metrics

Metric	Value
Validation Loss	0.5842
Validation Token Acc.	81.00%
Training Loss	0.6846
Training Runtime	9282s (~2.6h)
Samples/Second	5.21
Total Global Steps	1512

Hardware

GPUs: 2x Tesla T4
Platform: Kaggle / Cloud

License

Apache-2.0