dheeyantra/dhee-pravega
dheeyantra/dhee-pravega is a 5.1 billion parameter agentic tool-calling model developed by Dheeyantra, fine-tuned from google/gemma-4-E2B-it. It excels in complex tool selection scenarios across 18 languages, featuring enhanced accuracy when presented with multiple tool choices and supporting language switching. This model is specifically designed for applications requiring robust multilingual tool interaction and function calling.
Loading preview...
Dhee-Pravega v3: Multilingual Agentic Tool-Calling Model
dheeyantra/dhee-pravega v3 is a 5.1 billion parameter model, fine-tuned from google/gemma-4-E2B-it, specifically designed for agentic tool-calling across 18 languages. This version significantly improves tool selection accuracy in complex scenarios where multiple tools are available.
Key Capabilities and Improvements
- Enhanced Tool Selection Accuracy: Unlike previous versions trained on single-tool scenarios, v3 is trained on a diverse corpus where 68.8% of turns offer a genuine choice among several tools, and 12.3% correctly make no call at all. This leads to superior performance in real-world agentic applications.
- Vast Tool Schema Diversity: The training corpus for v3 includes 16,037 distinct tool schemas, a substantial increase from the 1,071 schemas seen by earlier versions, enabling broader tool integration.
- Multilingual Features: Introduces
change_languagefunctionality and translation prompts, expanding its utility in multilingual environments. - Performance Metrics: Achieves a 0.9847 rate for emitting expected tool calls and 0.9653 for correct tool selection on a held-out test set of 720 conversations across 18 languages, demonstrating strong reliability in challenging multi-tool contexts.
- Training Details: Utilizes QLoRA on the text decoder, trained on 181,191 conversations over 11,472 steps, continuing from the v2 adapter.
Ideal Use Cases
- Multilingual Agents: Developing AI agents that need to interact with tools and users in multiple languages.
- Complex Tool Orchestration: Applications requiring the model to intelligently choose from a wide array of available tools based on user intent.
- Function Calling: Scenarios where precise and reliable function calling is critical, especially when ambiguity or multiple options are present.
Note: This model is released under the Dhee Research-Only Licence v1.0, permitting academic research, evaluation, benchmarking, red-teaming, teaching, and personal experimentation. Commercial use requires a separate license from Dheeyantra Research Labs.