MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202641099275 A) filed by Tekframeworks Consulting Private Limited on August 17, 2026, for Corpus-State-Aware Adaptive Dataset Construction System And Method For Language Model Training.
Inventors include Venkata Dakshina Murthy Kolluru; Maheshkumar Duvvarapu; Vinay Chowdary; Venkata Dakshina Murthy Duddu; John Arul Prakash Sekar; and Arul Vignesh Arasan.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: CORPUS-STATE-AWARE ADAPTIVE DATASET CONSTRUCTION SYSTEM AND METHOD FOR LANGUAGE MODEL TRAINING Abstract of the Invention The present invention relates to a computer-implemented system and method for adaptive construction of training datasets for language models. Candidate data units are evaluated using training-utility signals including category coverage, novelty, source authority, model-derived confidence, and expected training impact. A corpus-state representation comprising category distribution, category imbalance, resource constraints, token-budget limitations, and training objectives is maintained. An adaptive decision engine jointly evaluates utility values and corpus-state information to determine dataset actions including retaining, excluding, relabelling, segmenting, or escalating candidate data units. The invention employs state-aware non- monotonic selection behaviour, enabling retention of data associated with underrepresented categories and de-prioritization of data associated with overrepresen
Disclaimer: Curated by HT Syndication.
