New research details Arkios, a 1.04B-parameter English-Nepali LLM with a Devanagari-aware tokenizer, and IndicDetect, a benchmark for AI-generated text detection in Hindi, Telugu, and Tamil. Another study explores cross-lingual transfer for Tulu legal comprehension.
New research introduces L3Cube-IndicQuest v2 for factual knowledge across 20 Indic languages, details Sarvam-105B's prompt injection safety, and proposes a Benchmark Maturity Index for Indian foundation models.
New SraVaani-1.0 ASR model, built on FastConformer, supports 65 Indian languages and dialects, trained on 31,255 hours of speech from the VAANI corpus, addressing a critical gap in inclusive speech recognition.
SurakshaEval is a new human-written safety benchmark for ten Indian languages and English, addressing cultural nuances and identifying recurring LLM safety failures.