The latest model significantly improves software building and migration capabilities in low-resource programming languages such as COBOL, Fortran, and Visual Basic. Despite scarce training data, these languages underpin critical global infrastructure (e.g., banking systems), and this enhancement aims to help teams maintain and modernize legacy code. Users can test via the API and provide feedback to shape future training.
Low-resource Language
低资源语言
Languages that lack large-scale annotated text data, making it difficult to directly train high-performance models.
Curated Intelligence
OpenAI announced significant improvements for legacy programming languages like COBOL and Fortran. GPT-6 Astra achieved an 87.4% average pass rate on MultiLangBench, up from 38% for GPT-5.6 Terra. On the COBOL migration benchmark, GPT-6.1 Sol reached 97.9% accuracy with high reasoning effort, addressing critical infrastructure modernization needs.
Latest AI models demonstrate enhanced capabilities for building and migrating software in low-resource languages such as COBOL, Fortran, and Visual Basic. This advancement targets the maintenance of critical infrastructure by offering API access for developer testing to refine future model training.
NVIDIA and AI4Bharat have released a massive 2.8 trillion-token Indic pretraining dataset covering native scripts for 22 languages and romanized data for twelve. The initiative also open-sources a reusable GPU-accelerated pipeline for translation, quality assessment, and repair, addressing critical data bottlenecks in training large language models for low-resource languages.
The Technology Innovation Institute (TII) has released Falcon ASR, an open-source speech recognition model designed to address the low-resource challenges of Arabic ASR. Built upon the Falcon LLM architecture, it is optimized for Arabic dialects and code-switching scenarios, providing high-precision, customizable local voice interaction infrastructure for the Middle East.
Google Research introduces WAXAL, a large-scale open speech dataset covering 27 Sub-Saharan African languages with ASR and TTS data. Licensed permissively, it aims to bridge the digital divide by empowering local ecosystems to build robust voice technologies for over 100 million speakers.
The article title introduces 'FilBench', a benchmark designed to test Large Language Models' understanding and generation of Filipino. Due to missing body content, specific evaluation methods, dataset composition, or experimental results are unavailable.
Hugging Face partners with IISc and ARTPARK to globally open-source the 'Vaani' dataset. As India's most diverse open-source multimodal and multilingual resource, it aims to lower barriers for building AI models tailored to Indian languages.
Hugging Face and Inception AI released AraGen, the first generative task benchmark and leaderboard for Arabic large language models. The initiative introduces 3C3H, a novel evaluation metric for Natural Language Generation (NLG), addressing critical gaps in robust assessment for low-resource languages. The team plans to extend this framework to other under-represented languages and invites community collaboration to maintain dynamic model rankings.
The Icelandic government plans to leverage GPT-4 technology to preserve and revitalize the Icelandic language. This initiative aims to use large models to process low-resource language data to prevent linguistic extinction. Currently, only an intention statement is disclosed, lacking specific implementation details and technical parameters.