kilian-group/KBevo-Qwen3-1.7B-SFT

KBevo-Qwen3-1.7B-SFT: Structured Knowledge and Reasoning

This model is a 2 billion parameter Qwen3-based language model, developed by kilian-group, specifically fine-tuned for multi-hop question answering. It implements the KBevo framework, which teaches the model to answer complex questions by first building a structured knowledge base and then reasoning over it. This approach involves two distinct, yet jointly trained, phases: knowledge-base construction and question answering through structured KB retrieval.

Key Capabilities

• Two-Phase Reasoning: Extracts (entity, relation, value) triples from text to form a knowledge base, then performs typed lookups for question answering.
• Grounded Answers: Answers are grounded in retrieved triples, reducing reliance on parametric memory alone.
• Supervised Fine-Tuning (SFT): This specific checkpoint represents the SFT stage of the KBevo framework, establishing the base extraction and lookup behavior.
• Research Foundation: Serves as the starting point for further optimization, such as the GRPO (Gradient-based Reinforcement Policy Optimization) training pipeline, and is detailed in the paper "Co-Evolving Structured Knowledge and Reasoning in Language Models" (COLM 2026).

Good For

• Reproducing Research: Ideal for researchers looking to reproduce the SFT-only KBevo behavior or to replicate the GRPO training pipeline described in the associated paper.
• Structured QA Development: Useful for developing and evaluating systems that require explicit knowledge extraction and reasoning over structured data.
• Multi-Hop Question Answering: Particularly suited for tasks like HotpotQA, where answers require synthesizing information from multiple sources.