zendesk-support-qwen2.5-7b-lora

Zendesk Support — Qwen2.5-7B-Instruct LoRA Adapter

A LoRA adapter fine-tuned on Qwen2.5-7B-Instruct for Thai customer support conversations exported from Zendesk (Sunshine Conversations).

Model Details

Training Details

Training Loss

• Initial training loss: 2.18
• Final training loss: 0.90
• Eval loss (epoch 1): 1.21 → 1.01
• Eval loss (epoch 2): 0.94 → 0.90

Usage

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-7B-Instruct",
    load_in_4bit=True,
)

from peft import PeftModel
model = PeftModel.from_pretrained(model, "usurachai/zendesk-support-qwen2.5-7b-lora")

# Inference
messages = [
    {"role": "system", "content": "You are a helpful, polite customer support agent for a Thai company. Respond in Thai."},
    {"role": "user", "content": "สั่งของไปแล้วยังไม่ได้รับเลยครับ"},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Infrastructure

Dataset

The dataset was exported from Zendesk via Sunshine Conversations, cleaned and deduplicated with:

• PII redaction (phone numbers, emails)
• Canned message deduplication
• Thai filler particle cleaning (ครับ/ค่ะ/ฮะ/etc)
• Sentence-level deduplication
• 1,808 training + 206 validation examples

License

Same as Qwen2.5 — Apache 2.0