Overview

This model, blackhao0426/pref-extractor-qwen3-0.6b-full-sft, is a specialized preference extraction model built upon the Qwen3-0.6B architecture. Its primary function is to analyze conversational turns between a user and an AI assistant and output structured JSON representing user preferences. Each preference is defined by a (condition, action, confidence) tuple, allowing for granular and contextual understanding of user needs.

Key Capabilities

• Structured Preference Extraction: Converts natural language dialogue into machine-readable JSON objects, detailing user preferences as condition-action rules.
• Lightweight Design: Based on the 0.6B parameter Qwen3 model, offering efficient deployment and inference.
• High Recall: Achieves 97.5% recall in preference extraction, intentionally over-extracting to ensure no preferences are missed, with a downstream filtering mechanism expected.
• Core Component of VARS: Integral to the VARS (Vector-Adapted Retrieval Scoring) framework, which aims to create personalized LLM assistants by modeling user preferences from interactions.

Performance & Training

Evaluated on a held-out test set, the model demonstrates 99.7% JSON validity and 97.5% recall, with a precision of 37.7%. It was trained for 1 epoch on the blackhao0426/user-preference-564k dataset, comprising 564,000 examples, using LLaMA-Factory with a learning rate of 2e-05.

Good For

• Developers building personalized LLM agents that need to adapt their behavior based on explicit or implicit user preferences.
• Applications requiring structured extraction of user interaction patterns for further processing or analysis.
• Research into user preference modeling and retrieval-augmented interaction in conversational AI systems.