benchmark

README

Benchmark (Model Quality / Ground Truth)

Click here for the results of pretraining 16 language models on different tokenizers including TokenMonster, GPT-2 & tiktoken p50k_base.

Benchmark Performance (Tokenization)

Click here for an interactive benchmark for all pretrained vocabularies, plus LLaMa Tokenizer, GPT2 Tokenizer, and tiktoken.

Click here for a chart comparing optimization mode, chr/tok and vocab size for all TokenMonster vocabularies.

Benchmark Performance (Speed)

Single threaded performance

	Range
TokenMonster	5.0 - 13 MB / second
tiktoken	4.0 - 7.0 MB / second
Hugging Face	0.2 - 0.5 MB / second

TokenMonster vs. Hugging Face (LLaMa Tokenizer)

The LLaMa Tokenizer running on Hugging Face Transformers used 35210511 tokens to tokenize "instruct" and 38383671 for "scifi", at an average speed of 0.29 MB/s. The TokenMonster import of LLaMa Tokenizer (with the exact same vocabulary), used 35083428 tokens for "instruct" and 38124152 for scifi (0.5% less tokens) and ran at an average speed of 12.5 MB/s.

TokenMonster vs. tiktoken (GPT2 Tokenizer)

Comparing to tiktoken's own benchmarks, they are claiming 6.2 MB/s with GPT2 Tokenizer. I got the same performance for GPT2 on tiktoken. TokenMonsters' import of GPT2 Tokenizer performed at 13.3 MB/s on "instruct" and 11.3 MB/s on "the_pile".

Notes on performance

The Python TokenMonster implementation calls the Go implementation for tokenization, so they are the same speed, but Python has overhead due to serialization and deserialization. The real-world performance of TokenMonster trained vocabularies, i.e. not gpt2 or llama (which are simpler vocabularies) is closer to the 5 - 7 MB/s range, which is comparable with the tiktoken performance. It's worth mentioning that tiktoken is tokenizing greedily whilst TokenMonster achieves comparable or better speed whilst calculating up to 6 branches at any point in time.

Benchmark Datasets

For a fair test, the benchmarks were performed on datasets that the TokenMonster vocabularies had not previously seen.

the_pile is the test dataset from The Pile, with the text extracting using extract_text_from_jsonl_parquet.py. Size is 1.3 GB.

github is this random file (1.7 GB direct download) from urls.txt from Red Pajama. It was also extracted using extract_text_from_jsonl_parquet.py. It represents code. Extracted size is 1.5 GB.

instruct is a bunch of chat & instruct finetunes from WizardLM's alpaca_evol_instruct_70k.json. This was used as-is and respresents chatbot conversational text. Size is 137 MB.

scifi is Scifi Stories Text Corpus. I thought sci-fi would be a good test for the fiction tokenizing capability because the training datasets don't contain much sci-fi. Size is 149 MB.

github and the_pile were further process with onlyvalidlatin.go to remove any invalid UTF-8 and non-Latin characters (e.g. Chinese). I made this decision because all of the pretrained vocabularies were trained with -only-latin and -only-valid parameters, hence they must use single byte tokens to tokenize any non-Latin characters. Because github and the_pile contained a lot of non-Latin script, whilst scifi and instruct did not, this would otherwise skew the benchmarks.

.