Large Language Model Integration

The project integrates with Large Language Models (LLMs) to help generate and optimize trading strategies from natural language descriptions.

Two main techniques are used:

Fine-tuning

Fine-tuning involves taking a pre-trained model and training it further on a custom dataset. This allows the model to specialize in interpreting and generating trading strategy code.

The dataset consists of natural language strategy descriptions paired with their code implementations. It is created using the training_data module.

To learn more about fine-tuning, check the Fine-tuning section.

Few-shot prompting

Few-shot prompting improves the performance of a language model by providing a few illustrative examples along with the user’s query. In this project, few-shot prompting is used to guide the LLM in generating precise outputs by showing it prior examples of strategy descriptions and their corresponding DSL representations.

The examples are carefully constructed to cover typical strategy structures and edge cases, helping the model understand the expected format and logic without needing additional fine-tuning.

To learn more about Few-shot prompting, check the Few-shot prompting section.

---

LLama

In this project we use the LLama model for LLM integration. We chose this model because of its open-source nature and its ability to be fine-tuned on custom datasets.

Running locally

To run the LLM locally we used the Ollama framework. Because of the hardware limitations, we used 1B and 3B Llama models which can be fine-tuned on a single GPU or Neural Engine in Apple Silicon which was our case.

Generating strategies

For generating strategies we use two approaches:

• First is to generate whole strategy object from natural language description. For this approach we use:

  • Fine-tuning: The 3B model
  • Few-shot prompting: The 3B model

  This model is more capable of generating complex code which is needed for generating whole strategy object. The model is trained on a dataset of natural language descriptions and their corresponding strategy objects.

• Second is to generate only parameters of the strategy listed in the Strategy module and then combine them into the strategy object. For this approach we use:

  • Fine-tuning:
    • 1B model for all of the parameters except the buy_condition and sell_condition.
    • 3B model for the buy_condition and sell_condition parameters. This is because these parameters are more complex and require a more capable model to generate them correctly.
  • Few-shot prompting:
    • In this case we use the 3B model for all of the parameters. This is because Few-shot prompting needs more context to generate the parameters correctly and the 3B model is more capable of generating complex code.