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, derived from Qwen/Qwen3-1.7B. It undergoes a unique two-stage training process to specialize its capabilities.

Key Capabilities & Training

• Topology-Guided Knowledge Distillation: In its first stage, the model was fine-tuned from Qwen/Qwen3-1.7B using Qwen/Qwen3-8B as a teacher model. This stage focused on transferring mathematical reasoning behavior, with a particular emphasis on derivation, verification, and high-discrepancy reasoning transitions, utilizing a dataset of advanced calculus problems.
• Agentic Specialization: The second stage involved LoRA specialization and weight merging, using the NousResearch/hermes-function-calling-v1 dataset. This stage specifically trained the model for assistant and tool-call outputs, making it proficient in structured function calling.
• Mathematical Replay: During agentic specialization, mathematical replay was incorporated to mitigate catastrophic forgetting of the reasoning capabilities learned in the first stage.

Intended Uses

• Mathematical and Technical Reasoning: Excels in tasks requiring logical derivation and problem-solving.
• Structured Function Calling: Capable of generating tool calls for external execution.
• Agent-Loop Research: Suitable for experiments involving tool selection and agentic workflows.
• Research on Topology-Aware Distillation: Provides a foundation for further study into this training methodology.

While the model can generate tool calls, it requires an external runtime to parse, execute, and integrate tool results back into the conversation flow. No formal benchmark results are claimed, and users should conduct their own evaluations for production use.