CentificAIResearch/RL_Adversarial_Dialogue_Agent

RLEaaS · RL-ADA Arena

Demo dashboard for co-evolutionary adversarial training of a banking Dialogue Agent (DA, 3B) against an adversarial Customer Agent (CA, 7B).

Modes

• —Fixture mode (default): replays recorded real arena trajectories. Runs on a free CPU Space — no GPU, no model servers needed.
• —Live mode: set the Space variable RLADA_LIVE=1. The Live Sim tab then calls real vLLM endpoints (DA :8000, CA :8002). Only works if those endpoints are reachable from the Space.

Tabs

• —🎥 Live Sim — watch a CA↔DA episode (fixture replay or live), with a live per-turn rule-based reward rubric.
• —📼 Replay — step through saved arena episodes (local only).
• —📊 Match History — DA win rate across co-evolutionary cycles (local only).
• —🧪 Verifiers & Rewards — the reward components and win-condition verifiers.
• —🏋️ Training Arena — orchestration control plane (placeholder).