muradil211/ToolWeave_stage2
muradil211/ToolWeave_stage2 is a 4 billion parameter language model based on the Qwen3-4B-Instruct architecture, developed by Muradil-mamat-211. This model is the second stage in the ToolWeave curriculum, focusing on Progress-Reward Learning to optimize multi-turn task completion through correct tool use. It is specifically trained to turn correct tool execution into measurable multi-turn task progress within complex environments, building on foundational tool competence. With a 32768 token context length, it aims for reliable task completion in tool-augmented scenarios.
Loading preview...
ToolWeave Stage 2: Progress-Reward Learning
muradil211/ToolWeave_stage2 is a 4 billion parameter model built upon the Qwen3-4B-Instruct base, representing the second stage of the ToolWeave curriculum. Its core focus is Progress-Reward Learning, where the model is trained to optimize actual progress in multi-turn environmental interactions by rewarding correct tool use.
Key Capabilities & Training Focus
- Multi-turn Task Completion: Specifically designed to improve performance on complex, multi-step tasks requiring sequential tool interactions.
- Progress-Reward Signal: Utilizes a fixed-denominator Progress Reward during training, directly linking correct tool execution to measurable task advancement.
- Curriculum-Based Development: Builds directly on the tool competence established in ToolWeave Stage 1, moving towards reliable task completion rather than just correct tool formatting.
- Context Length: Supports a substantial context length of 32768 tokens, crucial for understanding and executing multi-turn tool-use scenarios.
Evaluation Highlights
Internal validation on val_400_combined (400 examples across Base, Long Context, Missing Function, and Missing Parameter categories) shows an overall Progress Score of 0.4567. The model demonstrates higher progress on 'Base' tasks (0.6027) and maintains reasonable terminal coverage across categories. This evaluation uses a Progress Reward score, which is not directly comparable to Stage 1's format-gate score, indicating a shift in evaluation focus towards actual task progress.
Usage
Developers can integrate this model using the Hugging Face transformers library. Tool-use inference requires adherence to the model's function schemas and the Qwen3-compatible tool-call format. This release is a selected checkpoint and not the final ToolWeave Stage 3 model.