Call vLLM once with a list of prompts. Returns a list of completions in the same order.

Call OpenAI once per prompt, in parallel using threads. Returns a list of completions in the same order as prompts.

return the size of the input grid

return the size of the output grid

Parses a string representation of a 2D array into a NumPy ndarray. Parameters: - array_str (str): A string representation of a 2D array,

Extract integer sub-scores and final score from GPT output.

Read tasks from a JSON file.

Read tasks from a JSON file.

Read tasks from a file

Read tasks from a single JSON file with optional solutions.

Reset all LoRA B parameters to zero and LoRA A parameters to their initial values.

Run the incremental loop over items and return mean/std matrices.

Ask GPT-4.1 to score the completion on four criteria (1-5) and a Final Score. Returns a dict with keys: length, diversity, quality, correctne

Send one request to the TTT server and wait (≤10 min) for its reply. On timeout we recreate the REQ socket and retry, up to `max_retries`.

Send a single request to the inner-loop server and return the JSON reply.