Amid instances of biasness, incorrect and misleading information from the artificial intelligence (AI) large language models (LLMs), enterprise AI company Cloudera says tools such as fine-tuning studios can play a key role in reducing such hallucinations.
Fine-tuning studios such as those being provided by Cloudera, allow developers and companies to train their general-purpose Al models on domain-specific data. Training the model on more relevant and accurate information, makes it better at generating responses, thereby minimising chances of hallucinations.
“We believe fine-tuning is going to be important to reduce hallucinations. As part of our AMP (accelerators for machine learning projects) offering, we are providing a fine-tuning kind of service that customers can take advantage of,” Venkat Rajaji, senior vice president of product management at Cloudera, told FE.
According to Rajaji, along with fine-tuning, developers of AI models are also focusing on Retrieval-Augmented Generation (RAG) applications to provide additional context, additional data points, etc, to reduce the hallucinations.
RAG applications are required to improve the accuracy and relevance of AI-generated responses, particularly for complex queries that need precise information. It involves two steps – retrieval and generation of information. Retrieval step ensures that the response is up-to-date and directly related to the query asked. The generation step makes the output more fluid and human-like.
“It’s about context specific and providing the specific context for the model to be able to reduce hallucinations and provide the right answers. There are still some elements that are needed to be determined to reduce hallucinations,” Rajaji said.
Cloudera helps enterprises process data on any public or private cloud, into valuable and trusted insights. Last week, Cloudera introduced AMPs for enterprises to help with rapid application development. This includes a pre-built framework by Cloudera on chatbots, fine-tuning of models, prompt engineering, among others, and organisations can then build on top of that.
“We have 35 AMPs. If you add the number of AMPs associated with Hugging Face (AI community), it comes up into thousands of accelerated projects that we have as well. So, this volume really gives customers the ability to really accelerate application development leveraging our applications,” Rajaji said.
The company said it is taking an open ethos approach to the AMPs. “We register them to the AMP catalogue and then it becomes available to everybody who’s using Cloudera Machine Learning (CML),” Rajaji said.
When asked if these automated processes reduce the need for software developers going forward, Rajaji said, “AI is going to help them accelerate and make their jobs easier, make them more productive and probably churnout innovation faster”.
From India, Cloudera expects its business from India to double in FY25, similar to the performance in the last financial year. Globally, the company surpassed the $1 billion revenue mark in FY24.
In India, the company counts National Stock Exchange (NSE), PhonePe, Bharti Airtel, Axis Bank, among others as its customers.
“India is a very high growth market for us. We are seeing quite a bit of demand,” Rajai said, adding that financial services and banking is a strong industry for the company.
Besides, Cloudera is also seeing good traction from the government sector. Basis the demand in the Indian market, Cloudera is looking to increase its sales and the go-to-market team in a bid to tap more customers and grow revenue.