Sr Aivante

Sanskrit-322M-Base is a 322-million-parameter Sanskrit language model trained from scratch with its own 32,000-token SentencePiece Unigram tokenizer. It is intended for Sanskrit text completion, generation experiments, and further task-specific fine-tuning. This is a base pretrained model: it predicts the continuation of text. It has not been instruction-tuned or trained as a conversational assistant. It uses a Llama-compatible decoder architecture; its weights were initialized and trained independently, rather than adapted from a pretrained Llama checkpoint. Published by sraivante under the Apache License 2.0. The tokenizer uses identity normalization, byte fallback, and SentencePiece…

An offline English and Hinglish command interpreter for home devices and registered industrial/robotics devices. It converts one instruction into a catalog-validated JSON intent. The package includes a trained 87,181-byte action model, a 1,002,414-device SQLite catalog, a Python API, an HTTP service, a browser playground, Docker files, and reproducible training and evaluation data. The runtime was previously packaged locally as edge-command-docker-v3-ui. The system resolves device names through explicit identifiers, aliases and SQLite FTS5 search. It prioritizes action patterns and capability checks, with a stored count-based unigram/bigram classifier for the fallback scoring path. The…

A 97.6M-parameter language model trained from scratch, then fine-tuned for short, polite, everyday English conversation. It is a small, open research and learning model: you can read every line of its training code, run it on a laptop CPU, and see exactly where a model of this size is good and where it fails. - HellaSwag 33.6% (accnorm). That is above GPT-2 small (~30%) and below SmolLM2-135M (43.1%), which saw about 250x more training text. - The custom PyTorch code is included. The model does not use transformers; see How to use. This is QuickTalk run 8. "LMLM97M1" is its published name. 1. Learning and teaching how LLMs work. It is a complete, small, from-scratch GPT, with its training…

Versioned training, device-catalog and diagnostic data for Publisher: sraivante. Release date: 2026-09-26. Dataset version: v1.0.0. Instructions in English and Romanized Hindi/Hinglish map to a single structured device intent. The release preserves the raw command source, the recovered model training sample, exact diagnostic inputs, all predictions and the logical device catalog. There is no separate fine-tuning dataset for this count-based model. The test split is a diagnostic, not a held-out generalization benchmark. Training and diagnostic record-content overlap is also 82. There is no separate held-out validation set and no independently collected real-world test set. Configurations…