Vocabulary extension of the Mistral 7B tokenizer with Greek tokens for lower costs and faster inference (1.52 vs. 6.80 tokens/word for Greek)
8192 context length
We extend the pretraining of Mistral 7B with added proficiency for the Greek language, by utilizing a large corpus consisting of approximately 55 billion tokens.
This corpus includes 43.3 billion monolingual Greek tokens, constructed from publicly available resources. Additionaly, to mitigate catastrophic forgetting and ensure that the model has bilingual capabilities, we use additional sub-corpora with monolingual English texts (10.5 billion tokens) and Greek-English parallel data (600 million tokens).
This corpus has been processed, filtered, and deduplicated to ensure data quality (a detailed description of our data processing pipeline will be published in our upcoming paper) and is outlined below:
In the table below, we list the number and percentage of tokens for pretraining Meltemi 7B v1.5 (the respective values for Meltemi 7B v1 in parentheses).
Sub-corpus
# Tokens
Percentage
Greek
43,383,244,502 (28,555,902,360)
79.5% (72.0%)
English
10,538,413,259 (10,478,414,033)
19.3% (26.4%)
Parallel
633,816,023 (633,816,023)
1.2% (1.6%)
Total
54,555,473,784 (39,668,132,416)
100%
Meltemi 7B v1.5 was trained for less than 2/3rds of the training steps of Meltemi 7B v1.
Usage
Please make sure that the BOS token is always included in the tokenized prompts. This might not be the default setting in all evaluation or fine-tuning frameworks.
Evaluation
The evaluation suite we created includes 6 test sets and has been implemented based on a fork of the lighteval framework.
*** The differences in the Meltemi 7B v1 scores compared to the ones here can be attributed to a different -and better optimized- evaluation setup for Greek, i.e., lighteval vs. lm-eval-harness. ***
An existing benchmark for question answering in Greek (Belebele)
A novel benchmark created by the ILSP team for medical question answering based on the medical exams of DOATAP (Medical MCQA).
Our evaluation is performed in a few-shot setting, consistent with the settings in the Open LLM leaderboard.
We can see that our new training procedure enhances performance across all Greek test sets by a +16.3% average improvement compared to Mistral 7B. The results for the Greek test sets are shown in the following table:
Medical MCQA EL (15-shot)
Belebele EL (5-shot)
HellaSwag EL (10-shot)
ARC-Challenge EL (25-shot)
TruthfulQA MC2 EL (0-shot)
MMLU EL (5-shot)
Average
Mistral 7B
29.8%
45.0%
36.5%
27.1%
45.8%
35%
36.5%
Meltemi 7B v1
46.3%
68.5%
63.3%
43.6%
44.6%
42.4%
51.4%
Meltemi 7B v1.5
48.1%
68.6%
65.7%
47.1%
45.1%
42.4%
52.8%
Ethical Considerations
This model has been aligned with human preferences, but might generate misleading, harmful, and toxic content.
Acknowledgements
The ILSP team utilized Amazon’s cloud computing services, which were made available via GRNET under the OCRE Cloud framework, providing Amazon Web Services for the Greek Academic and Research Community.
Citation
@misc{voukoutis2024meltemiopenlargelanguage,
title={Meltemi: The first open Large Language Model for Greek},
author={Leon Voukoutis and Dimitris Roussis and Georgios Paraskevopoulos and Sokratis Sofianopoulos and Prokopis Prokopidis and Vassilis Papavasileiou and Athanasios Katsamanis and Stelios Piperidis and Vassilis Katsouros},
year={2024},
eprint={2407.20743},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.20743},
}
