alea-institute/kl3m-004-128k-uncased-mlm

kl3m-004-128k-uncased tokenizer

NOTE: This is the same vocabulary as kl3m-004-128k-uncased, but packaged within a RobertaProcessing post_processor class to provide special token handling without loading a custom tokenizer class.

The kl3m-004-128k-uncased case-insensitive tokenizer is a domain-specific tokenizer trained on a stratified sample of nearly 4M documents across general, legal, and financial domains from the kl3m-data project, including American English, British English, Spanish, German, French, Italian, and other common EU languages.

This tokenizer is being used for the next generation of KL3M embedding and generative models.

Please see kl3m-001-32k and kl3m-003-64k for the first iteration of our research on domain-specific tokenization.

Note that we are providing both uncased and cased versions of the 128K tokenizer, unlike prior tokenizers, as this was required to achieve SotA in-domain performance for embedding models on legal and financial text.

Model Details

Summary

• Vocabulary: 131,072
• Tokenizer type: BPE
• Special token support: Both causal and masked language modeling
• Language(s) (NLP): Primarily English, Spanish, German, French, with a small percentage of other EU languages.
• Data Sources: See kl3m-data repository.
• Developed by: ALEA Institute.
• License: CC-BY 4.0

Model Description

The kl3m-004-128k-uncased tokenizer is a domain-specific tokenizer trained on ~1.5T tokens of financial and legal text from primarily-English sources.

This tokenizer is notable for a number of reasons:

Domain Specific

As part of our research on more efficient SLM training for the legal and financial domain, we trained a domain-specific tokenizer on a large corpus of financial and legal text. This tokenizer has not, for example, seen any common general pretrain sources like Wikipedia or Common Crawl.

Large Added Token Set

As part of our research on efficient and reliable extraction and generation, we inserted a large numer of deterministic "whole" tokens into the tokenizer, such as HTML tags like <span, common Markdown elements like # and ##, and legal enumerations like (a).

Note that the kl3m-004-128k-uncased tokenizer has added a number of additional citation formats that were not included in the kl3m-001-32k tokenizer. These were primarily sourced from empirical data and the Free Law Project's reporters-db, which were added to the tokenizer to improve model behavior related to legal citations.

See the get_custom_tokens method in kl3m_embeddings/training/kl3m_004/train_tokenizer.py for more details:

def get_custom_tokens(
    include_whitespace: bool = True,
    include_markdown: bool = True,
    include_html: bool = True,
    include_json: bool = True,
    include_xml: bool = True,
    include_years: bool = True,
    include_citations: bool = True,
    lowercase: bool = False,
) -> list[str]:

Space Preservation

Unlike kl3m-001-32k, we do not retain the space character as a token. This was done after adding additional legal citation tokens to the vocabulary, which reduced the number of issues related to space tokenization in legal text. This means that the kl3m-004-128k-uncased tokenizer uses substantially fewer tokens than kl3m-001-32k for most text.

Special Tokens for both Embedding and Generative Models

For both training and inference efficiency, we intended this tokenizer vocabulary to be usable for both embedding and generative models. As such, we included special tokens suitable for both causal and masked language modeling tasks.

• <|start|>: 0
• <|end|>: 1
• <|pad|>: 2
• <|unk|>: 3
• <|sep|>: 4
• <|cls|>: 5
• <|mask|>: 6

We also added a number of chat and instruction tokens that were not included in kl3m-001-32k, including:

• <|system|>: 7
• </|system|>: 8
• <|user|>: 9
• </|user|>: 10
• <|instruction|>: 11
• </|instruction|>: 12

These tokens are identical to those used in the kl3m-003-64k tokenizer.

Replication

The entire data collection and preprocesing pipeline is being made available, along with training data, as part of the ALEA Institute KL3M project.

The source code to used to train the tokenizer is available on GitHub at: https://github.com/alea-institute/kl3m-embedding-research

The data pipeline will be available on GitHub and S3 in the near future.

Uses

This tokenizer is intended to be used for English, Spanish, German, or French language text in professional contexts such as legal and financial documents.

Recommendations

In general, the kl3m-004-128k-uncased tokenizer is recommended over the original kl3m-001-32k tokenizer.

Original text:  The Comptroller of the Currency shall have the same authority with respect to functions transferred to the Comptroller of the Currency under the Enhancing Financial Institution Safety and Soundness Act of 2010 as was vested in the Director of the Office of Thrift Supervision on the transfer date, as defined in section 311 of that Act [12 U.S.C. 5411].

kl3m-003-64
-----------
Size:  67
Tokens:  ['The', ' Comptroller', ' of', ' the', ' Currency', ' shall', ' have', ' the', ' same', ' authority', ' with', ' respect', ' to', ' functions', ' transferred', ' to', ' the', ' Comptroller', ' of', ' the', ' Currency', ' under', ' the', ' Enh', 'ancing', ' Financial', ' Institution', ' Safety', ' and', ' ', 'Sound', 'ness', ' Act', ' of', ' 2010', ' as', ' was', ' vested', ' in', ' the', ' Director', ' of', ' the', ' Office', ' of', ' Thrift', ' Supervision', ' on', ' the', ' transfer', ' date', ',', ' as', ' defined', ' in', ' section', ' 311', ' of', ' that', ' Act', ' [', '12', ' ', 'U.S.C.', ' 54', '11', '].']
IDs:  [671, 13273, 295, 281, 25922, 735, 704, 281, 1913, 2451, 440, 1894, 312, 5860, 7264, 312, 281, 13273, 295, 281, 25922, 621, 281, 18926, 4406, 3195, 24448, 5617, 310, 233, 63589, 2130, 854, 295, 1611, 398, 725, 11978, 300, 281, 2827, 295, 281, 1767, 295, 44029, 37141, 395, 281, 3696, 1548, 24, 398, 3011, 300, 782, 6590, 295, 407, 854, 1327, 524, 233, 63761, 3789, 547, 8578]


kl3m-004-128k-uncased
---------------------
Size:  64
Tokens:  ['the', ' comptroller', ' of', ' the', ' currency', ' shall', ' have', ' the', ' same', ' authority', ' with', ' respect', ' to', ' functions', ' transferred', ' to', ' the', ' comptroller', ' of', ' the', ' currency', ' under', ' the', ' enhancing', ' financial', ' institution', ' safety', ' and', ' soundness', ' act', ' of', ' 2010', ' as', ' was', ' vested', ' in', ' the', ' director', ' of', ' the', ' office', ' of', ' thrift', ' supervision', ' on', ' the', ' transfer', ' date', ',', ' as', ' defined', ' in', ' section', ' 311', ' of', ' that', ' act', ' [', '12', ' ', 'u.s.c.', ' 54', '11', '].']
IDs: [536, 16356, 292, 281, 4272, 460, 628, 281, 1552, 1545, 397, 882, 309, 4378, 4032, 309, 281, 16356, 292, 281, 4272, 539, 281, 21164, 1271, 3843, 2737, 313, 35934, 638, 292, 2371, 363, 611, 5286, 298, 281, 2456, 292, 281, 1652, 292, 25900, 7290, 390, 281, 1397, 643, 24, 363, 1921, 298, 590, 12646, 292, 384, 638, 745, 629, 233, 128952, 3834, 571, 4442]  

How to Get Started with the Model

Use the code below to get started with the model.

from tokenizers import Tokenizer

tokenizer = Tokenizer.from_pretrained('alea-institute/kl3m-004-128k-uncased')

Citation

Tokenizer and dataset publications are pending.

Contact

For any questions, please contact ALEA Institute at [email protected] or create an issue on this repository or GitHub.