Moses AI Blog

We are excited to share our latest research on selective quantization techniques designed to minimize general quality loss during model compression for language model transport. This work addresses a critical challenge in deploying large language models: efficiently transferring models while maintaining their performance across diverse tasks.

Our selective quantization approach strategically identifies and preserves critical model components during the compression process. The key insight is that near quality loss free compression can be achieved by only compressing the weights that are not susceptible to precision reduction. By selectively protecting sensitive weights while aggressively compressing others, we achieve significantly reduced quality degradation compared to traditional uniform quantization methods.

This enables more efficient model distribution and deployment without sacrificing the versatility that makes language models valuable. The technique allows for substantial model size reduction while maintaining performance across diverse tasks.

Publication coming soon. Stay tuned for detailed technical insights and experimental results.

We introduce selective simulated quantization, a novel regularization technique specifically designed to improve task-specific performance in small language models with fewer than 2 billion parameters. This approach addresses the unique challenges faced when working with resource-constrained models that need to excel at specific tasks.

Through our experiments, we have achieved up to 10% improvements in task-specific perplexity using this technique. By simulating quantization effects during training and selectively applying them to specific model components, we enable small language models to develop more robust representations that generalize better to their target tasks.

These task-specific models can be tested on request. If you're interested in exploring how this technique could benefit your use case, please reach out via the contact form.

To tackle one of the core problems of large language models and to make Moses outputs more reliable and usable for business, Moses features enhanced hallucination controls as part of its chat widgets.

It thereby imposes output filters that check for hallucinated content. If hallucinated content is detected, the response will indicate that the agent cannot truthfully respond.

Under the hood of the Moses platform, large language models (LLM) are used to extract semantics of input texts as well as for text generation. At their core, these LLMs are neural networks, artificial replicas of structures that also can be found in the human brain.

Similar to a human mind, also neural networks can become biased, even become racist when trained with a dataset from which certain correlations that we deem discriminating can be inferred. While precautions are taken to reduce the risk of such bias, the nature of mathematical function fitting, which training of neural network resembles quite closely cannot exclude the risk of a LLM being discriminating.

To comply with the General Data Protection Regulation (GDPR), specifically with respect to the Risks to the Rights and Freedoms of Natural Persons it is mandatory to not discriminate based on origin, race, gender or other components of Personal Identifiable Information (PII).

To exclude the risk of discrimination through the models in place by Moses AI, we introduce the PII redaction widget. It can be placed anywhere in the virtual agents logic and redacts personal identifiable information (PII) in the question and replaces it by [PII] and overwrites the question. It is trained to detect PII data such as addresses, credit card numbers, CVV, names, social security numbers.

We are pleased to announce that our service at Moses AI is fully compliant with the European General Data Protection Regulation (GDPR). The GDPR, which went into effect in May 2018, is a comprehensive set of regulations aimed at protecting the personal data of individuals within the European Union (EU). It applies to any organization that processes the personal data of EU citizens, regardless of where the organization is located.

As part of our compliance efforts, we have conducted a thorough review of our data processing practices and have ensured that we are compliant with the GDPR. We have also implemented procedures for obtaining informed consent, as well as for handling requests for data access, rectification, or erasure.

We are committed to protecting the privacy and security of our users' data. Hence, we have ensured that our AI service is designed to process personal data, where personal data is applicable or required, in a secure and transparent manner, ensuring that individuals' rights are protected at all times. We will continue to monitor developments in data protection legislation, and if updates to our service are necessary, we will make them.

If you have any questions about our compliance with the GDPR or about our data processing practices, please do not hesitate to contact us at support@moses-ai.com

Moses provides a set of widgets specific for the Instagram platform and others that can be used in the social media context. Once you have Moses account you can get started. You can decide yourself if you want to plug the Moses virtual agent into your own application and use it via our HTTP API, or you can use our mobile application that automatically uses the agent logic you have built for it.

The rude language filter and the sensitive language widget can be connected to the Instagram Input widget, filtering comments and comment replies that contain sensitive or rude language.

Both the hide comment widget and the write reply widget generate executable HTTP GET requests. Instead of sending these, they are written to the admission queue. This queue allows for adding a human into the loop, who can review and modify response texts and hide actions.

In a push to make Moses virtual agents truly multi-modal, Moses features a Speech-to-text (STT) widget. It is capable of transcribing audio files (in wav or mp3 format) to text that is used in the downstream logic from the STT widget as the input question.

This feature allows for building virtual agents that can react to human speech. When combined with text-to-speech engines, the Moses API can be used to power conversational agents that do not require written input.

The Moses platform now features the PDF analysis widget. In context of a Moses virtual agents logic, it can be applied by first connecting a request PDF upload widget and then a PDF to text widget.

The Moses API will prompt for a PDF upload which is then analyzed using optical character recognition. The transcript is then used by the following widgets as the input query, opening it to the full suite of Moses widgets to further analyze the transcript and perform actions based on the results.