experimentlab

EXPERIMENT LAB presents Vigyan—a suite of specialized, open-weight Small and Medium Language Models (2B, 7B, and 32B) domain-pretrained on a 1B+ token curated STEM and technical literature corpus.

While generalist LLMs face high cloud costs, severe hallucination rates, and data privacy risks in specialized domains, Vigyan provides localized, high-precision reasoning directly on low-power edge hardware with zero cloud API dependencies.

Key Features & Benchmarks

• High-Precision STEM Reasoning: Achieves 54.3% accuracy on multi-step scientific reasoning and digital logic verification benchmarks.
• Biomedical & Chemical Accuracy: Reaches 91.3% SMILES retrosynthesis validity (+14.4% over frontier generalist models) and 99.2% stereochemistry preservation.
• Edge-Native & Privacy-First: Runs fully offline in ~4.8 GB quantized VRAM at 1/40th the inference compute cost of proprietary cloud APIs, guaranteeing total data sovereignty.

Live Links & Resources

• 🌐 Website:experimentlab.in
• 🤗 HuggingFace Checkpoints:olmo2-7b-biomed-chem
• 💻 GitHub Codebase:shreyansh001boy-tech