Occiglot-7B-ES-EN

A polyglot language model for the Occident.

Occiglot-7B-ES-EN is a generative language model with 7B parameters for Spanish and English and trained by the Occiglot Research Collective. It is based on Mistral-7B-v0.1 and trained on 112B tokens of additional multilingual and code data with a block size of 8,192 tokens per sample. Note that the model is a general-purpose base model and was not instruction-fine-tuned nor optimized for chat or other applications. We make an instruction tuned variant available as occiglot-7b-es-en-instruct

This is the first release of an ongoing open research project for multilingual language models. If you want to train a model for your own language or are working on evaluations, please contact us or join our Discord server. We are open for collaborations!

Model details

• Continued-pretraining from: Mistral-7B-v0.1
• Model type: Causal decoder-only transformer language model
• Languages: English, Spanish, and code.
• License: Apache 2.0
• Compute resources: HessianAI's 42
• Contributors: Manuel Brack, Patrick Schramowski, Pedro Ortiz, Malte Ostendorff, Fabio Barth, Georg Rehm, Kristian Kersting
• Research labs: Occiglot with support from SAINT and SLT
• Contact: Discord

How to use

You can use this model directly with a pipeline for text generation. Since the generation relies on some randomness, we set a seed for reproducibility:

>>> from transformers import pipeline, set_seed
>>> generator = pipeline('text-generation', model='occiglot/occiglot-7b-es-en')
>>> set_seed(42)
>>> generator("Hola, soy una modelo lingüística", max_length=40, num_return_sequences=1)
[{'generated_text': 'Hola, soy una modelo lingüística que puede ayudarte a traducir textos entre español e inglés. Si me envías un texto en español'}]

Dataset

The training data is the respective subset of the data used for occiglot-7b-eu5, i.e. Spanish plus English and Code.

The data distribution by language (estimated) is as follows:

• English: ~34%
• Code: ~13%
• Spanish: ~52%

The training data was prepared using lm-datasets. The exact data configuration is here.

Training settings

• Continual pre-training on 128 x A100-80GB on HessianAI's 42.
• Framework: Determined
• Precision: bf16
• Optimizer: AdamW (lr: 0.00001, warmup_steps: 420)
• Global batch size: 512 (with 8192 blocksize) split over 128 GPUs
• Cosine Annealing with Warmup

Tokenizer

Tokenizer is unchanged from Mistral-7B-v0.1.

Evaluation

Preliminary evaluation results can be found below. Please note that the non-English results are based on partially machine-translated datasets and English prompts (Belebele and Okapi framework) and thus should be interpreted with caution, e.g., biased towards English model performance. Currently, we are working on more suitable benchmarks for Spanish, French, German, and Italian.

All 5 Languages

English

Spanish

Acknowledgements

The model training was supported by a compute grant at the 42 supercomputer which is a central component in the development of hessian AI, the AI Innovation Lab (funded by the Hessian Ministry of Higher Education, Research and the Art (HMWK) & the Hessian Ministry of the Interior, for Security and Homeland Security (HMinD)) and the AI Service Centers (funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK)). The curation of the training data is partially funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK) through the project OpenGPT-X (project no. 68GX21007D).

License

Apache 2.0

See also

• https://huggingface.co/collections/occiglot/occiglot-eu5-7b-v01-65dbed502a6348b052695e01