reaperdoesntknow/DualMind-TKD-Agentic-1.7B

DualMind-TKD-Agentic-1.7B Overview

DualMind-TKD-Agentic-1.7B is a 1.7 billion parameter model developed by reaperdoesntknow, built upon the Qwen/Qwen3-1.7B base. Its unique capabilities stem from a two-stage training lineage designed to enhance both mathematical reasoning and agentic function-calling.

Key Capabilities

• Topology-Guided Knowledge Distillation: The first stage involved distilling mathematical reasoning from Qwen/Qwen3-8B using a dataset of advanced calculus problems. This process focused on transferring derivation, verification, and high-discrepancy reasoning transitions.
• Agentic and Function-Calling Specialization: The second stage fine-tuned the model for agentic behavior and structured function calling, utilizing the NousResearch/hermes-function-calling-v1 dataset. Mathematical replay was integrated to mitigate catastrophic forgetting during this stage.
• Standalone Model: The repository contains the merged standalone model, eliminating the need for a separate PEFT adapter for inference.

Intended Uses

• Mathematical and Technical Reasoning: Optimized for tasks requiring complex derivations and logical thought.
• Structured Function Calling: Capable of generating tool calls for external execution within an agentic workflow.
• Agent-Loop Research: Suitable for experiments in tool selection and agent behavior.
• Research on Topology-Aware Distillation: Provides a foundation for further study into this advanced training methodology.