kl3m-004-char-8k-cased

The kl3m-004-char-8k-cased case-sensitive tokenizer is a domain-specific character-based tokenizer trained on a stratified sample of nearly 2M 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 uses the standard Byte-Pair Encoding (BPE) tokenizer from tokenizers/transformers, but modifies the training process to restrict the vocabulary to tokens that are at most 3 characters long. Models trained with this tokenizer should be able to handle a number of use cases that are otherwise difficult to handle with standard tokenizers, such as low-resource spell-checking, OCR correction, whitespace normalization, and other tasks that require a high degree of character-level granularity.

Model Details

Summary

• Vocabulary: 8,192 tokens
• Tokenizer type: BPE with 1-3 character tokens
• 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

For more information about the kl3m-004 tokenizers, see the kl3m-004-128k-cased tokenizer.

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.

This specific tokenizer was trained using the following command:

PYTHONPATH=. poetry run python3 \
  kl3m_tokenizers/tokenizers/kl3m_004/train_char_tokenizer.py \
  --min_frequency 1000 \
  --vocab_size 8192 \
  --pad2 \
  --max_chars 3 \
  sample.20241223173012.jsonl.gz \
  ./kl3m-004-char-8k-cased/

Training tokenizer.
[00:32:57] Pre-processing sequences       █████████████████████████████████████████ 1849343  /        0
[00:33:18] Pre-processing sequences       █████████████████████████████████████████ 0        /        0
[00:00:21] Tokenize words                 █████████████████████████████████████████ 20286360 / 20286360
[00:00:57] Count pairs                    █████████████████████████████████████████ 20286360 / 20286360
[00:11:11] Compute merges                 █████████████████████████████████████████ 7844     /     7844
Adding power-of-2 padding tokens.
Padded vocab to 8192 tokens.
Special tokens: 13
Power-of-2 pad tokens: 13
Final vocab size: 8192
Training time: 2759.02 seconds
Output path: kl3m-004-char-8k-cased

Uses

This tokenizer is intended to be used for English, Spanish, German, or French language tasks where character-level details are important, such as OCR correction, spell-checking, or tasks where word boundaries are not well-defined.

For a standard BPE "word" tokenizer with a larger vocabulary size, consider using the kl3m-004-128k-cased or kl3m-004-128k-uncased tokenizers.

Recommendations

The kl3m-004-char-8k-cased tokenizer may be particularly useful when a smaller vocabulary size is needed and character-level details are important. For larger vocabularies and better handling of word-level and subword-level patterns, consider using the kl3m-004-128k-cased or kl3m-004-128k-uncased tokenizers.

4k and 16k character tokenizers are also available for tasks that require more character-level detail, although these tokenizers may be less efficient or require more memory to use.

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-char-8k-cased')

Citation

Tokenizer and dataset publications are pending.

Contact

For any questions, please contact ALEA Institute at hello@aleainstitute.ai or create an issue on this repository or GitHub.