Model Overview

EnvScaler-Qwen3-1.7B is a 1.7 billion parameter language model built upon the Qwen3 architecture, specifically designed for tool-interactive agent tasks. Developed by XXHStudyHard, this model leverages the EnvScaler framework to enhance its capabilities in environments requiring tool use.

Training Methodology

The model was trained using a two-stage approach:

• Supervised Fine-Tuning (SFT): Initial training involved 9,022 trajectories from agent-environment interactions, utilizing data from EnvScaler-SFT-Traj-9K across 4,684 SFT scenarios and 141 synthesized environments.
• Reinforcement Learning (RL): Further refinement was conducted using 2,550 RL scenarios and 50 synthesized environments, based on the ROLL framework. This stage allows the model to learn from reinforcement signals, optimizing its performance in complex interactive tasks.

Key Capabilities

• Tool-Enhanced Interaction: Optimized for scenarios where the language model needs to interact with external tools or environments.
• Agentic Performance: Designed to function effectively as an agent, learning from both demonstrations and reinforcement.
• Specialized Training: Benefits from a unique training regimen focused on agent-environment interactions, distinguishing it from general-purpose LLMs.

Use Cases

This model is particularly suited for applications requiring an LLM to act as an agent, interact with various tools, and navigate complex environments. Developers can integrate it with the EnvScaler project for full functionality in tool-interactive settings.