bmonikraj/qwen3-8b-search-plan-act
bmonikraj/qwen3-8b-search-plan-act is an 8 billion parameter Qwen3 model, fine-tuned using GRPO on the `search_plan_act` task. This model is specifically optimized for multi-turn tool-calling scenarios, enabling it to search for facts, manipulate records, and achieve goals through iterative tool use. It demonstrates significant performance improvements in complex reasoning and tool-use benchmarks like MuSiQue and BALROG BabyAI, making it suitable for applications requiring advanced planning and action capabilities.
Loading preview...
Model Overview
This model is a full-weight checkpoint of Qwen/Qwen3-8B, fine-tuned using GRPO (Generative Reinforcement Learning with Policy Optimization) on the search_plan_act task. It is designed for complex, multi-turn interactions where the model needs to perform actions like searching for information, reading/updating records, and linking entities to achieve a specific goal.
Key Capabilities
- Advanced Tool-Calling: Trained with TRL's multi-turn
environment_factorytool-calling API, enabling sophisticated interaction with external tools. - Procedural Task Execution: Excels at
search_plan_acttasks, which involve iterative tool calls until a goal is met or the task is abandoned. - Reward-Decomposed Training: Utilizes a reward system that considers outcome, grounding, and stop-behavior components, preventing reward-hacking strategies.
- Improved Reasoning: Demonstrates substantial performance gains on benchmarks like MuSiQue (EM / F1: 0.108 / 0.191 pre-RL to 0.338 / 0.456 post-RL) and BALROG BabyAI (mean episode return: 0.000 pre-RL to 0.020 post-RL), indicating enhanced reasoning and planning abilities.
Good For
- Applications requiring models to perform sequential, goal-oriented actions using external tools.
- Tasks involving information retrieval, record management, and complex decision-making in a closed-book setting.
- Developing agents that can plan and execute multi-step strategies to solve problems.