--- license: apache-2.0 datasets: - PleIAs/common_corpus language: - en - fr - es - de - it - la - nl - pl ---
**Pleias-3b-Preview** is an early preview of a 3 billion parameters base model trained by [Pleias](https://huggingface.co/PleIAs) on [Common Corpus](https://huggingface.co/datasets/PleIAs/common_corpus). Pleias-3b-Preview was pretrained at Jean Zay (compute grant n°GC011015451) with support from [Etalab](https://www.etalab.gouv.fr/) Like all the base and specialized models from Pleias, Pleias-3b-Preview has only been trained on open data out of copyright (public domain) or under a permissible license. ## Description Pleias-3b-Preview is a transformer base model, entirely pretrained from scratch, using an architecture similar to Llama/GPT-Neox for easier deployment/inference. It includes the following features, that would apply to any responsibly trained variant: * Only trained on open data under a permissible license and in compliance with the European AI Act. By design, all Pleias model are unable to output copyrighted content. * Extensive multilingual support for main European languages. * A new tokenizer designed for enhanced document processing tasks and better multilingual support. * Extremely low level of toxicity and problematic content. Fully supported languages include English, French, Spanish, German, Italian, Dutch, Latin and Portuguese. ## Recommended use As a base model, Pleias-3b-Preview is only able to run continuation prompts. Text generation is currently able to support a range of creative writing tasks in multiple European languages. For more consistent results we recommend using a low or null temperature with a slight repetition penalty (1.1-1.2). Pleias-3b-Preview has been successfully adapted for continuous pretraining and full-fine-tuning on document processing tasks such as RAG, translation or OCR correction. Given the small size of the model we do not recommend fine-tuning methods based on LORA. ## Training Pleias-3b-Preview was fully pretrained at Jean Zay on 192 h100s for about 20 days (compute grant n°GC011015451) with support from [Etalab](https://www.etalab.gouv.fr/). Training code relied on Nanotron, the open source library from HuggingFace. We provide the complete settings as a yaml file as part of our release. Training schedule includes 518,000 steps (batch size 1,024) on a filtered and enhanced version of Common Corpus (1,086,324,736,000 tokens). Training Greenhouse Gas Emissions: Estimated total location-based greenhouse gas emissions were 16 tons CO2eq for training. ## Ethical Considerations pleias-3B-Preview model, like all large language models, carries inherent ethical risks that require careful consideration. Our approach to mitigating these risks begins at the data level, where we exclusively use vetted sources, deliberately excluding CommonCrawl. The primary challenge comes from our public domain dataset component, which contains historical texts that may reflect outdated social norms and potentially harmful language, particularly regarding minoritized groups. To address this, we implemented a systematic ethical filtering process using toxicity classifiers to identify extremely harmful content. We also employed synthetic rewriting techniques to transform mildly problematic passages while preserving the underlying informational value. This process significantly reduced potential societal harm without compromising the dataset's size or textual quality, resulting in notably low toxicity scores in benchmarks compared to other models. Despite these preventive measures, users should be aware that the model has not undergone additional safety alignment procedures and may still produce problematic outputs. The model's capabilities in generative AI tasks must be balanced against the risks of bias, misinformation propagation, and autonomous decision-making challenges. We explicitly prohibit any malicious utilization and emphasize the responsibility of users to implement appropriate safeguards. At Pleias, we continue to research and develop improved methods for creating safer and more equitable models and datasets. This includes ongoing work in toxicity reduction, bias mitigation, and the development of more sophisticated ethical filtering techniques. ## Acknowledgements This work would not have been possible without the substantial support from étalab. The training was conducted as part of the Grand Challenge of GENCI, aligned with the European strategy for establishing AI factories through the EuroHPC Joint Undertaking, aimed at supporting European startups and providing open-source models to the community. We express our gratitude to GENCI's Jean Zay supercomputer, France's AI flagship facility, which was instrumental in our model's training. The project benefited from the new NVIDIA H100 partition specifically dedicated to the French AI community. We appreciate the generous allocation of compute hours over five months and the invaluable technical expertise provided by IDRIS, EVIDEN, and NVIDIA (as well as its Inception program). We are deeply grateful to the Mozilla Foundation Local AI Program for their generous support. Finally, we acknowledge the significant contributions from the open science LLM community, particularly HuggingFace, Eleuther AI and Allen AI whose insights and cooperation have been invaluable to our work. ## Update Pleias-3b-Preview is currently released as an early preview. The model will undergo several more round of post-training to enhance reasoning capacities and fine-tunability as well as in anticipation of a generalist instruct version.