Husky-v1 is an open-source language agent designed for complex, multi-step reasoning tasks, including numerical, tabular, and knowledge-based challenges. It targets researchers and developers working on advanced AI reasoning capabilities, offering a structured approach to problem-solving through iterative tool use.

How It Works

Husky-v1 operates in two stages: action generation and action execution. It iteratively generates the next best action using its internal models and then executes that action via specialized tools, updating the solution state. This approach leverages a code generator, query generator, and math reasoner as expert models, enabling it to tackle diverse reasoning problems efficiently.

Quick Start & Requirements

• Install dependencies with pip install -r requirements.txt.
• Requires CUDA 11.8.
• Download models from the official HuggingFace collection.
• Inference requires running multiple shell scripts in parallel (run_husky_action.sh, run_husky_code.sh, etc.), ideally with access to five GPUs.
• Datasets need to be downloaded and placed in the dataset folder.
• Official documentation and HuggingFace models are available.

Highlighted Details

• Utilizes vLLM for optimized, batched inference.
• Supports parallel execution of multiple expert models for enhanced performance.
• Offers flexibility in choosing action generator models (7B, 13B) and datasets.
• Includes evaluation scripts (eval_husky.sh) for assessing performance.

Maintenance & Community

Contact is available via GitHub issues or email for support.

Licensing & Compatibility

The repository does not explicitly state a license in the provided README.

Limitations & Caveats

The system requires CUDA 11.8 and recommends five GPUs for optimal inference, posing a significant hardware barrier. The README does not specify the license, which could impact commercial use or integration into closed-source projects.