Orbit Support Drafter (DPO)
The Drafter agent of a LangGraph multi-agent support system for the fictional shop Orbit
Electronics. Base model TinyLlama/TinyLlama-1.1B-Chat-v1.0, LoRA-fine-tuned and merged. Stage: dpo
(sft = supervised on the response simulator; dpo = then aligned with human and critic
preferences via Direct Preference Optimisation).
Evaluation (held-out simulated tickets, full multi-agent graph)
| stage | temperature | first-draft approval | resolved | avg drafts |
|---|---|---|---|---|
| base | 0 | 0 | 0 | 3 |
| sft | 0 | 1 | 1 | 1 |
| sft | 0.7 | 1 | 1 | 1 |
| dpo | 0 | 1 | 1 | 1 |
| dpo | 0.7 | 1 | 1 | 1 |
Knowledge-base fingerprint: 034b279610ce. Demo model trained on synthetic data; it only
knows the toy policies of Orbit Electronics.
Preferences: {'synthetic': 24, 'critic': 14, 'gold': 2} (β=0.1, epochs=2).