Train & Evaluate Safety models

This is the Open Assistant Safety Folder and contains the following:

• Model training scripts
• Model inference scripts
• Data processing scripts

Mission Statment

Our mission at LAION-AI OpenAssistant safety team is to create a safety pipeline that is not only compatible with the OpenAssistant model and project but can also integrate with other systems outside of it. We are dedicated to making this pipeline modifiable and robust to accommodate the diverse preferences of our users.

We understand that our users come from different backgrounds and use various types of hardware. Therefore, we strive to make our safety pipeline accessible and able to run on consumer hardware, so everyone can benefit from its protective features.

Through our commitment to innovation and collaboration, we will continue to provide safety solutions that ensure the well-being of our users and the wider community.

Why create a safety pipeline?

Open source and extendable safety pipelines unfortunately do not exist on the same scale as those in ChatGPT and other commercial systems. To further research in implementable, accurate, and extendable safety pipelines, Open Assistant Safety Team will continue to push models and code to the public. Much research has been done in things like toxicity detection and bias mitigation in LLMs, however the implementation of such research in systems that use language models as conversational agents in production settings has largely gone undocumented. Furthermore, safety systems that interact with diverse communities of users must be able accommodate user preferences. This is paramount in introducing LLM based systems all over the world. We hope that our work will generate more research in this field, and allow others to create safe LLM based systems.

Training

• Set training configuration using config.yaml

python model_training/t5_trainer.py