metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
tags:
- grpo
- rl
- support-ticket
- lora
- peft
- trl
Support Ticket GRPO Agent
Fine-tuned Qwen/Qwen2.5-0.5B-Instruct using GRPO (Group Relative Policy Optimization) + LoRA on a multi-step support ticket environment.
Training Setup
- Algorithm: GRPO via
trl.GRPOTrainer+ LoRA (PEFT) - Base model: Qwen/Qwen2.5-0.5B-Instruct
- Dataset: 1000 prompts over 50 support tickets
- Environment: algocore-support-ticket-env
- Group size G: 2
- KL beta: 0.04
- Final loss: 0.0008
Results
| Task | Before | After | Delta |
|---|---|---|---|
| Task 1 (Classify) | 0.667 | 1.000 | +0.333 |
| Task 2 (Action) | 0.117 | 0.450 | +0.333 |
| Task 3 (Full Resolve) | 0.083 | 0.258 | +0.175 |
| Overall | 0.289 | 0.569 | +0.280 |
