Revochat Responder (SFT, full)
The Responder agent of Revochat, a LangGraph multi-agent live-chat system for the fictional shop
ReVonance. It writes Revo's chat messages; a deterministic Session manager decides what to say.
Base model TinyLlama/TinyLlama-1.1B-Chat-v1.0, fine-tuned with full fine-tuning.
Stage: sft (sft = supervised on gold chat turns from the session simulator).
Evaluation (simulated chat sessions: 3 simulator LLMs x 10 personas, full multi-agent graph)
| variant | first-pass approval | chat goal success | CSAT |
|---|---|---|---|
| base | 0.013 | 0.908 | 4.139 |
| lora | 0.997 | 0.9 | 4.119 |
| full | 0.998 | 0.889 | 4.117 |
Knowledge-base fingerprint: f9cdb07b5df6. Demo model trained on synthetic data; it only
knows the toy policies of ReVonance.