Custom Generative AI Solutions with Watsonx.ai

Generative AI Effectively

A recent IBV study found that 64% of CEOs are pressured to speed generative AI deployment, and 60% lack a uniform, enterprise-wide process.

Watsonx, an AI and data platform, may help enterprises use foundation models and expedite generative AI adoption.

The newly launched features and capabilities of watsonx.ai include new general-purpose and code-generation foundation models, more open-source model options, and more data and tuning options to increase the business impact of generative AI. These improvements were guided by IBM’s strategic belief that AI should be open, trustworthy, focused, and empowering.

IBM-developed, data-driven foundation models for businesses

Business executives using generative AI require model flexibility and choice. They also require safe access to business-relevant models to speed up value and insights. IBM’s watsonx.ai studio offers a variety of language and code foundation models of various sizes and architectures to assist clients provide performance, speed, and efficiency.

Atsushi Hasegawa, Honda R&D Chief Engineer, believes watsonx.ai is a compelling solution in an environment where system integration and software connectivity are crucial. Its inherent flexibility, rapid deployment, and strong information security make it appealing.”

Watsonx.ai launched with the Slate series of encoder-only models for corporate NLP. We are pleased to offer Granite, our first IBM-developed generative foundation model. A decoder-only design makes the Granite model family ideal for summarization, content creation, retrieval-augmented generation, categorization, and insight extraction.

All Granite foundation models were trained using IBM-curated enterprise datasets. The Granite family of models was trained on enterprise-relevant datasets from five domains—internet, academic, code, legal, and finance—scrubbed for objectionable information and benchmarked against internal and external models to obtain deeper domain expertise. This technique reduces risks so model results may be safely released using watsonx.data and watsonx.governance (coming soon).

Initial IBM Research evaluations and testing across 11 financial tasks show that Granite-13B models can outperform much larger models by training them with high-quality finance data. Financial tasks assessed include sentiment scoring stock and earnings call transcripts, categorizing news headlines, extracting credit risk ratings, summarizing financial long-form content, and answering financial or insurance queries.

Adding openness to IBM AI models

Data provenance, testing, safety, and performance factors are missing from many AI models. Uncertainties can hinder generative AI adoption for many enterprises and organizations, especially in highly regulated industries.

IBM shares the Granite model training data sources today.

1. Common Crawl
2. Webhose
3. GitHub Clean
4. Arxiv
5. USPTO
6. Pub Med Central
7. SEC Filings
8. Free Law
9. Wikimedia
10. Stack Exchange
11. DeepMind Mathematics
12. Project Gutenberg (PG-19)
13. OpenWeb Text
14. HackerNews

IBM bases AI development on trust and openness. IBM will protect clients against third-party IP claims against IBM-developed foundation models as a testament to its rigorous development and testing. IBM does not force clients to indemnify IBM for using IBM-developed models, unlike some other Large Language Model vendors. IBM does not cap its IP indemnity liability for IBM-developed models, keeping with its indemnification policy.

We encourage clients to adapt IBM models for downstream activities as they utilize them to generate distinctive AI assets. Clients may ethically leverage their company data to improve model output accuracy and get a competitive edge through fast engineering and tuning.

Helping companies appropriately utilize third-party models

With dozens of open-source big language models, it’s hard to know where to start and which model to use. However, selecting the “right” LLM among hundreds of open-source models needs careful consideration of cost-performance considerations. Since many LLMs are unpredictable, AI ethics and governance should be included in model design, training, tweaking, testing, and outputs.

We established a foundation model library in watsonx.ai for clients and partners because one model won’t do. Starting with 5 handpicked open-source models from Hugging Face, we chose them after thorough technical, licensing, and performance analyses and identifying their optimal use cases.

This month, we introduced Meta’s 70 billion parameter model Llama 2-chat to watsonx.ai studio as an open-source LLM model. To converse and generate code, use Llama 2. It is pretrained with public web data and fine-tuned with human input reinforcement learning. Llama 2 improves virtual agent and chat applications for commercial and research use.

Watsonx.ai now offers BigCode’s StarCoder LLM. The model can explain and answer general code inquiries in natural language using permissively licensed GitHub data as a technical helper. It can also autocomplete, alter, and explain code fragments in natural language.

Third-party model users in watsonx.ai can enable AI guardrails to automatically eliminate objectionable language from input prompts and output.

Reduce model-training risk with synthetic data

Traditional data anonymization can generate flaws that adversely affect outputs and forecasts. By using computer simulation or algorithms to produce synthetic data, companies may fill data gaps and avoid the danger of revealing sensitive data.

Watsonx.ai’s synthetic data generator lets enterprises produce pre-labeled tabular data that retains their enterprise data’s statistical features. Instead of using extended data-collection periods to capture actual data’s enormous variance, this data may be utilized to tweak AI models faster or increase their accuracy by introducing more diversity into datasets. Building and testing models with synthetic data can help firms close data gaps and launch new AI products faster.

Empowering business use cases with fast tweaking

Tuning Studio at watsonx.ai helps business users customize foundation models for Q&A, content production, named entity recognition, insight extraction, summarization, and classification.

The initial Tuning Studio version will offer quick tuning. Advanced prompt tweaking in watsonx.ai (based on 100 to 1,000 samples) lets enterprises adjust foundation models to their own data. Prompt-tuning lets a corporation with minimal data adjust a big model to a small goal, reducing compute and energy cost without retraining an AI model.

Business AI advancement and assistance

IBM Watsonx is a business-focused AI and data platform that helps more people expand and accelerate AI’s impact using trustworthy data. The watsonx architecture allows new business-targeted foundation models like IBM Research’s and third-party models like those on the Hugging Face open-source platform to be seamlessly integrated as AI technologies advance, while providing critical governance guardrails with watsonx.governance.

Watsonx is one of IBM’s generative AI platforms. Over 1,000 IBM Consulting specialists can help customers tune and operationalize models for particular business use cases using generative AI.

Source