Open LLAMA 13B released, trained on 1T tokens

> Please note that it is advised to avoid using the Hugging Face fast tokenizer for now, as we’ve observed that the auto-converted fast tokenizer sometimes gives incorrect tokenizations.

If you are correct, maybe it'll work for languages not reliant on whitespace, paired with a code formatter, but we're definitely coming to expect more from these models these days.

> For current version of OpenLLaMA models, our tokenizer is trained to merge multiple empty spaces into one before tokenization, similar to T5 tokenizer. Because of this, our tokenizer will not work with code generation tasks (e.g. HumanEval) since code involves many empty spaces. We are planning to open source long context models trained on more code data. Stay tuned.

Sounds like they're working on it.

e.g. 31 spaces would be 16 + 8 + 4 + 2 + 1 = 5 tokens; 15 spaces would be 8 + 4 + 2 + 1, etc; but for a lot of code the spaces are an even multiple (such as when a tab = 4 spaces or tab = 2 spaces) so even fewer would be required.

If you took all of the Python code and queried how many spaces are used, I would wager most lines have four spaces, a few have eight, and even fewer have more than that.

If the Python code were written with four spaces = tab, then it wouldn't have cost any more tokens than squashing the repeated spaces (since a "four-space" token consumes the same as a "one-space" token).