ersaasyhar/d3-edge-indo-legal

Fine-tuned decision model based on vllm-sr/d3-edge (0.6B) specialized for Indonesian statutory tree navigation trajectories (root_to_bab and bab_to_pasal).

Model Details

• Base Architecture: Qwen3.5 0.6B with 255-way linear readout (readout.safetensors).
• Domain: Indonesian legal tree navigation (Undang-Undang, Peraturan Pemerintah).
• Optimization: 4-Epoch Elbow Method (Optimal Knee Point: Epoch 4).
• Validation Accuracy: 38.00% (Baseline: 38.00%).
• Validation Loss: 2.1307 (Baseline: 2.1307).
• Training Trajectories: ersaasyhar/d3-indo-legal-navigation.

Elbow Method Convergence Curve

Usage with D3Navigator

from src.tree_reasoning.d3_navigator import D3Navigator

navigator = D3Navigator(model_name="ersaasyhar/d3-edge-indo-legal")
results = navigator.retrieve("Apa syarat pendirian perseroan terbatas?", top_k=5)
for chunk in results:
    print(f"{chunk.doc_id} - {chunk.title} (p. {chunk.page_number})")