RLLM Training & Chat with Code¶

Null Terminal includes a powerful local engine for both "Chat with Code" (RAG) and "Scratch Training" (creating new models).

🧠 Chat with Code (RAG)¶

Index your local codebase to allow the AI to answer questions about your project.

Usage¶

1. Navigate to your project root.
2. Build the index:

   /index build
   

3. Ask questions: Just toggle AI mode (Ctrl+Space) and ask:

   "How does the authentication system work?" "Refactor the check_permissions function in utils.py"

The system uses semantic search to find relevant code snippets and inserts them into the context window.

🏋️ RLLM Scratch Training¶

Train your own small language models (like LLaMA or custom architectures) directly on your hardware.

1. The Training Screen¶

Access the training UI via the command palette (Ctrl+P) -> "Open Training Dashboard" or by typing /train.

2. Configuration¶

The dashboard allows you to configure text-based training parameters:

• Model Architecture: Define layers, hidden size, heads, etc. (e.g., LLaDA config).
• Dataset: Path to your local training data (streaming JSON/Text supported).
• Hyperparameters: Learning rate, batch size, context length.
• Optimization: Flash Attention, BF16 mixed precision, Activation Checkpointing.

3. Monitoring¶

Real-time metrics are displayed in the TUI: - Loss Curves: Training and validation loss. - Throughput: Tokens per second. - Hardware Stats: GPU/CPU usage and VRAM consumption.

4. Output¶

Trained checkpoints are saved to ~/.null/models/checkpoints. You can load these back into Null Terminal using the Custom Provider interface.

[!WARNING] Training requires significant hardware resources. Ensure you have an NVIDIA GPU with CUDA support for best performance.