Overview

Klingspor/Qwen3-4B-SFT is a specialized language model derived from Qwen3-4B, fine-tuned for the classic game of 20 Questions. Its primary function is to serve as a Questioner, strategically asking yes-or-no questions to identify a hidden common English noun.

Key Capabilities

• 20 Questions Game Play: Excels at asking deductive questions to narrow down possibilities in the 20 Questions game.
• Reinforcement Learning Initialization: Designed to be a strong starting point for further reinforcement learning (RL) training, specifically for models like StarPO and CIA, as detailed in the paper "Intrinsic Credit Assignment for Long Horizon Interaction."
• Interactive Agent Research: Supports research into multi-turn interactive language agents.

Training Details

The model underwent supervised fine-tuning (SFT) on successful, filtered single-turn trajectories. The training data comprised 341 words from the COCA (Corpus of Contemporary American English) word list, ensuring no overlap with subsequent RL training or test sets. A Qwen3-14B model acted as the judge or oracle during this process.

Intended Use Cases

• Developing and testing agents for the 20 Questions game.
• Providing a robust SFT checkpoint for advanced RL-based training experiments.
• Exploring and advancing research in interactive language agent design and long-horizon interaction.