PEFT
Safetensors
trl
sft
customer-support
ticket-routing
qwen3

Qwen3 0.6B Bitext Ticket Router SFT

LoRA SFT adapter for routing customer-support tickets to the top 12 intents by frequency from the first 50000 rows of bitext/Bitext-customer-support-llm-chatbot-training-dataset.

The training prompt enforces strict JSON-only output:

system: You are a customer-support ticket router. Return only one strict JSON object...
user: <ticket text>
assistant: {"intent":"<label>","confidence":0.99,"reason":"<short reason>"}

Eval

Held-out eval examples: 1000

{
  "accuracy": 0.999,
  "valid_json_rate": 1.0,
  "schema_pass_rate": 1.0,
  "accuracy_on_schema": 0.999,
  "num_eval_examples": 1000,
  "num_train_examples": 10793,
  "top_intents": [
    "check_invoice",
    "complaint",
    "contact_customer_service",
    "edit_account",
    "switch_account",
    "check_payment_methods",
    "contact_human_agent",
    "delivery_period",
    "get_invoice",
    "newsletter_subscription",
    "payment_issue",
    "registration_problems"
  ],
  "base_model": "Qwen/Qwen3-0.6B",
  "dataset": "bitext/Bitext-customer-support-llm-chatbot-training-dataset",
  "max_steps": 800,
  "learning_rate": 0.0002,
  "warmup_ratio": 0.03,
  "max_length": 512
}
Downloads last month
66
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for OldEngine/qwen3-0.6b-bitext-ticket-router-sft

Finetuned
Qwen/Qwen3-0.6B
Adapter
(432)
this model

Dataset used to train OldEngine/qwen3-0.6b-bitext-ticket-router-sft