yannabadie/sage-topology-orchestrator

SAGE Topology Orchestrator

Base: nvidia/Nemotron-Orchestrator-8B (original weights, tool-calling preserved) Training: DAPO via verl 0.7.1 on 2x H100 NVL 94GB Format: JSON tool-calls (<tool_call> native format, NOT YAML) Framework: YGN-SAGE

Why JSON, not YAML

Nemotron-Orchestrator-8B was GRPO-trained by NVIDIA to emit structured JSON tool calls. Previous YAML training caused 91% malformation. JSON uses the model's native format.

Training Status

Not yet started — dataset conversion and script preparation in progress (March 30, 2026).

SAGE Modules as Tools

The model learns to orchestrate YGN-SAGE modules:

• add_node(role, model_tier, prompt) — create topology node
• add_edge(from_idx, to_idx) — connect nodes
• set_reasoning(text) — explain topology choice
• checkpoint(node_idx, fallback_tier) — mark adaptation point
• upgrade(node_idx, new_tier) — upgrade model at checkpoint (Phase C)
• continue() — proceed to next node (Phase C)
• reroute() — abort and restart (Phase C)

Previous Model

See yannabadie/sage-topology-policy-v2 for the YAML-based training history (Phase A step 1050, DAPO step 300).