Qwen2.5-7B-SupportRAG-LoRA (Milestone 2 - DA3 Aligned)

Fine-tuned QLoRA adapter for Qwen2.5-7B-Instruct, specialized for multi-tool agentic customer support reasoning using ChatML syntax (<|im_start|> and <|im_end|>).

Model Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Method: QLoRA 4-bit SFT (LoRA rank=16, alpha=32)
• Training Duration: 507 Steps (~4h 15m on Google Colab T4)
• Primary Tasks: Grounded knowledge retrieval, order tracking verification, and policy-compliant human escalation.
• Tools Supported:
  • knowledge_base_search(query, brand)
  • check_order_status(order_id)
  • escalate_to_human(reason, brand, urgency)

Quickstart

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

base_model_id = "Qwen/Qwen2.5-7B-Instruct"
adapter_id = "LoneVertex/Qwen2.5-7B-SupportRAG-LoRA"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id)

prompt = """<|im_start|>system
You are a customer support assistant with tool access.<|im_end|>
<|im_start|>user
[AmazonHelp] How do I track my package #ORD-9821?<|im_end|>
<|im_start|>assistant
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

System Architecture

This adapter forms the core offline-trained reasoning engine of the SupportRAG Chatbot production system.