Add im_start / im_end in cl100k_base

[spolu]

Not really directly useful given the Chat API...

But triangulating:

• https://github.com/openai/openai-cookbook/blob/main/examples/How_to_format_inputs_to_ChatGPT_models.ipynb
• https://github.com/openai/tiktoken/commit/ec7c121e385bf1675312c6c33734de6b392890c4#diff-0d973848bd229418209db2c46c86167000845592ca6b98fad215c21c317bc494R9

We know they exist.

[zyxue]

what does im mean in im_start/im_end?

[spolu]

They are the special tokens used in the OpenAI Chat format as it gets translated and presented to the model.

[youkaichao]

@zyxue It seems to be "input message". Check here.

[microsoftbuild]

@spolu

It's possible that the APi side uses and extended tokenizer like: https://github.com/openai/tiktoken/tree/main#extending-tiktoken