Released our pre-print OpenBezoar: small, cost-effective and open models trained on mixes of instruction data on arXiv