Steve Dept (he/him/his)’s Post

How do I Prompt an AI-Powered TMS? - Learn how AI-powered translation management systems analyze and improve translations without prompts. Discover the capabilities of large language models in translation processes. https://hubs.la/Q02LqvLV0

How do I Prompt an AI-Powered TMS? - Learn how AI-powered translation management systems analyze and improve translations without prompts. Discover the capabilities of large language models in translation processes. https://hubs.la/Q02LqvvX0

How do I Prompt an AI-Powered TMS? - Learn how AI-powered translation management systems analyze and improve translations without prompts. Discover the capabilities of large language models in translation processes. https://hubs.la/Q02LqzqK0

It feels nice seeing that my models on Hugging Face 🤗 have surpassed 500,000 downloads since their upload. Here’s a bit more about them: - SMaLL100: A highly compact, massively multilingual machine translation model that supports 100 languages while delivering performance comparable to M2M-100 12B: https://lnkd.in/evwJUE8G - RQUGE: A model designed for evaluating large language models (LLMs). It was initially used to assess the question generation task in this project. The model was inspired by QAFactEval work: https://lnkd.in/eutzaD9f

Llama 3.1 is getting great results. The secret sauce? It learned more languages... 😏 #MultilingualAI #GlobalAI Llama 3.0: "To prepare for upcoming multilingual use cases, over 5% of the Llama 3 pretraining dataset consists of high-quality non-English data that covers over 30 languages. However, we do not expect the same level of performance in these languages as in English." Llama 3.1: “Data mix summary. Our final data mix contains roughly 50% of tokens corresponding to general knowledge, 25% of mathematical and reasoning tokens, 17% code tokens, and 8% multilingual tokens.” “We use a vocabulary with 128K tokens. Our token vocabulary combines 100K tokens from the tiktoken tokenizer with 28K additional tokens to better support non-English languages. Compared to the Llama 2 tokenizer, our new tokenizer improves compression rates on a sample of English data from 3.17 to 3.94 characters per token. This enables the model to “read” more text for the same amount of training compute. We also found that adding 28K tokens from select non-English languages improved both compression ratios and downstream performance, with no impact on English tokenization.” Llama 3.1 Paper:

Machine translation (MT) has revolutionized the way we communicate across language barriers, making it a vital tool in technical, commercial, government, and internet domains. Despite its imperfections, MT’s widespread use underscores its significance in today’s globalized world. This blog post delves into the intricacies of machine translation, examining its applications, benefits, limitations, and future prospects.

Imagine a world where language barriers do not exist; a tool so intuitive that it can understand the subtleties of every dialect and the jargon of any industry. While we’re not quite there yet, advancements in Large Language Models (LLMs) are bringing us closer to this vision.

Low switching costs are a key factor supporting the commoditization of Large Language Models (LLMs). The simplicity of transitioning from one LLM to another is largely due to the use of a common language (English) for queries. This uniformity allows for minimal cost when switching. #llm #opensource https://lnkd.in/gpKM44av

Now that the cost of using large language models (LLMs) is plummeting, this topic is a shot as it gets for language solution providers: At what pace and for what tasks do we use large language models (LLMs) in upstream linguistic quality assurance, translation, adaptation to conform to local usage, downstream linguistic quality control? Should we have multiple connectors and APIs to call up neural machine translation (NMT) engines as well as LLMs from our computer-assisted translation tools (CAT tools), or should we concentrate on a small subset of best-of-class engines? For what projects is customization worth it? Is translation in context by LLMs consistently (un)reliable? Are there good metrics to help us make informed decisions? Translated ventures some answers in their white paper. Here is their pitch: https://lnkd.in/ee2EDxHS

IBM Client Engineering & Consulting engaged in a pilot with the City of Helsinki introducing watsonx.ai for summarization and translation to compare the efficacy of LLM models versus Watson Language Translator. By applying watsonx.ai LLM-based translations in City of Helsinki chatbots to summarize and translate foreign language questions into Finnish, we demonstrated enhanced customer care services by improving intent recognition through question summarization and translation. Leveraging watsonx Assistant as the intent classifier and using translations from watsonx.ai, we showcased: COMET scores .90 average with minimal prompt engineering “The pilot developed by IBM demonstrated that the idea of utilizing LLMs as a translation technology has very promising and real potential. And although more planning and testing is needed, we are confident we can together design a solution that is both safe to use, and provides actual value to the customer and us as the service provider. Our longstanding close collaboration with IBM specialists facilitates the development of solutions also based on entirely new technologies as part of our existing services.” Janne Kantsila, Tech Lead, product owner. City of Helsinki That’s the power of IBM. What can we do for your organization? #ibm #clientengineering #innovation watsonx #ai #showdonttell https://lnkd.in/e6eupesQ