qvac/genesis-i-model

QVAC Genesis I Pretrained Model Overview

The qvac/genesis-i-model is a 1.7 billion parameter decoder-only transformer, developed by Qvac by Tether, and represents the first publicly released open-source pretrained model specifically built on a rigorously validated synthetic dataset for education. It was trained from scratch on approximately 40 billion tokens from Tether's QVAC Genesis I, the largest synthetic educational dataset, utilizing BF16 mixed precision and a 4,096-token context window.

Key Capabilities

• Multi-Domain Educational Coverage: Inherits curriculum-aligned knowledge across mathematics, physics, biology, and medicine due to its training on the QVAC Genesis I dataset.
• Superior Benchmark Performance: Consistently outperforms baselines in reasoning tasks, knowledge assessments, and subject-specific question answering, leveraging its specialized training data.
• Base Pre-trained Model: Provided as a base model, it is intended for further continual pre-training or fine-tuning for specific educational applications.
• Qwen3-family Architecture: Built with a Qwen3-family 1.7B-parameter decoder-only transformer architecture.

Good for

• Continual Pre-training (CPT) or Fine-tuning (SFT): Ideal for specializing in STEM-focused tutoring, QA systems, and curriculum support.
• Benchmarking: Useful for evaluating reasoning and subject-specific QA performance in educational contexts.
• Research: Supports research into synthetic dataset-driven LLM training methodologies.
• General Language Modeling: Capable of next-token prediction, continuation, summarization, and drafting.