Horde Agent

LLM Training Agent for SFT/GRPO/DPO workflows.

Structure

• agent/ - Core orchestration logic
• tools/ - Individual tool implementations
• workflows/ - Workflow definitions
• ui/ - Streamlit interface
• config/ - Configuration files
• scripts/ - Executable scripts

Quick Start

Local Development (CPU)

# Create CPU virtual environment
python -m venv venv-cpu
source venv-cpu/bin/activate  # On Windows: venv-cpu\Scripts\activate

# Install CPU-only dependencies
pip install -r requirements-cpu.txt

# Run pipeline
python -m agent.main \
  --data_path "your-dataset" \
  --out_dir "./output/run1" \
  --max_iters 3

# Or use UI
streamlit run ui/app.py

Production (GPU)

# Create GPU virtual environment
python -m venv venv-gpu
source venv-gpu/bin/activate  # On Windows: venv-gpu\Scripts\activate

# Install GPU dependencies
pip install -r requirements.txt

# Run pipeline
python -m agent.main \
  --data_path "your-dataset" \
  --out_dir "./output/run1" \
  --max_iters 3

Supported Training Methods

• SFT (Supervised Fine-Tuning)
• GRPO (Group Relative Policy Optimization)
• DPO (Direct Preference Optimization)