Support Ticket Classification with LoRA
A LoRA fine-tuning experiment based on Qwen/Qwen2.5-0.5B-Instruct for support ticket queue and priority prediction.
Training
- Training examples: 18,998
- Validation examples: 2,375
- Training duration: 1 epoch, 1,188 steps
- Method: LoRA with PEFT
Preliminary evaluation
The quick evaluation used 100 validation examples:
- Queue accuracy: 43%
- Queue macro-F1: 0.3202
- Priority accuracy: 39%
- Priority macro-F1: 0.2957
- Valid JSON rate: 100%
These are preliminary results, not a full-validation benchmark. Baseline comparisons require identical evaluation examples.
Usage
Load the base model Qwen/Qwen2.5-0.5B-Instruct and then load this adapter with PEFT's PeftModel.from_pretrained.
Limitations
This is an experimental model. Verify predictions before using them for real customer support decisions.
- Downloads last month
- 10