Training Data Module

The training_data module is responsible for generating synthetic and structured datasets used to fine-tune language models for the purpose of understanding and generating trading strategies. It produces prompt-response pairs useful for supervised fine-tuning.

---

Module Structure

• training_data.py
  Entry point for generating complete training datasets, combining all individual components.

• prompt_builder.py
  Constructs natural language prompts from structured strategy definitions.

• condition_generator.py
  Randomly generates valid strategy entry/exit conditions.

• interval_generator.py
  Supplies typical time intervals defined here.

• stop_loss_take_profit_generator.py
  Generates stop-loss and take-profit configurations.

• capital_size_commission_generator.py
  Produces initial capita, order size and commission configurations.

• strategy_type_generator.py
  Defines whether the strategy is long-only, short-only, or both using PositionTypeEnum.

• start_end_date_generator.py
  Sets up historical date ranges for backtests.

• ticker_generator.py Randomly selects from a pool of ticker symbols. Tickers are defined in sp500.csv.

---

prompt_data/ Submodule

• condition_dicts.py
  Provides mappings of condition types and corresponding indicators.

• string_options.py
  Contains sets of predefined string values from which to randomly select to generate prompts.

• sp500.csv)
  A list of S&P 500 tickers used by the ticker_generator.

---

How It Works

The module is designed to be highly modular. You can: - Use each generator individually - Combine them via training_data.py to create full training examples - Set random seeds for reproducibility

---

What does it generate?

Vie the training_data.py module, you can generate a dataset of prompt-response pairs. Every time you call the generate_trading_data() function, it will create a new dataset with the specified number of examples. It always creates prompt that is a text in natural language and response to the prompt extracting from the prompt:

• Whole Strategy object
• Ticker
• Position Type
• Conditions
• Stop Loss
• Take Profit
• Start Date
• End Date
• Interval
• Period
• Initial Capital
• Order Size
• Trade Commissions

All of those datasets are saved as .jsonl files and used for fine-tuning the models.