atekrugis/distilroberta-base-v4-gold-router

DistilRoBERTa V4 Gold Router

This model is a high-precision intent router designed to classify user prompts into 6 distinct categories for LLM orchestration.

📊 Categories & Label Mapping

• —0: agentic (Weather, searches, real-time data)
• —1: coding (Programming tasks, debugging)
• —2: general (Chitchat, greetings, basic info)
• —3: math (Calculations, LaTeX, logic)
• —4: reasoning (Chain-of-thought, philosophy, deep analysis)
• —5: tool_calling (API structured requests)

🛠️ Training Details

• —Base Model: distilroberta-base
• —Dataset: V4 Gold Purified (approx. 46k rows)
• —Accuracy: 99.8% on validation set
• —Key Fixes: Resolved the "Reasoning Bias" found in V3 and stabilized against NaN gradient explosions.