Generic extractor for score/reason pair: - Try JSON parsing first - Fallback to regex patterns if JSON fails

Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): T

Call GPT with images/text and retry on failure.

This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

Initialization for cache.

Encode image to base64 string.

Send a chat completion request with prompt and optional images, with retry logic. Returns GPT response content or empty string on failure.

Send chat completion request with text+image inputs, retry up to 3 times Returns assistant's content or empty string on failure

Normalize the str to lower case and make them float numbers if possible.

Set random seeds for reproducibility

Do not add eos again if user already added it.

Ensure the value is divisible by the stride.

Ensure the value is divisible by the stride.

Ensure the value is divisible by the stride.

Ensure the value is divisible by the stride.

Ensure the value is divisible by the stride.

Calculate the instruction level accuracy for given Subject results

Create a logger that writes to a log file and stdout.

This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to the docstring of this me

Read an image file and return its base64-encoded string (or None on error).

Encode image file to base64 string.

Read image file and return base64-encoded string

Encode an image file to a base64 string.

Call GPT with 3 reference images and 1 predicted image in one shot.

Send reference frames + predicted frame to GPT and return its response. Frames are sent in numerical order with the predicted frame inserted