EurIPS

This release includes a fully working pipeline to reproduce the workshop paper experiments:

• SFT training of Llama 3.2 1B on Pubchem filtered (molgen40M)
• fine-tuning of SMI-TED for property prediction
• RLFT with GRPO on SMI-TED property predictors (multi-reward) with physics-informed equations of state
• pre-trained checkpoints available