search

IBM / molformer

Repository for MolFormer
Apache License 2.0
244 stars 42 forks source link

Inaccurate reproduction results on QM9 #15

Closed Park-ing-lot closed 11 months ago

Park-ing-lot commented 12 months ago

When I finetune the model from checkpoint "N-Step-Checkpoint_3_30000.ckpt", the reproduction results are much higher than them report in the paper.

For example, I got 1.4575 on u0 task, while the reported result on the same task is 0.3211. I have checked the finetune code several times, I cannot find issues.

Is there anyone who get similar results to the reported ones on any regression task?

hoffmansc commented 11 months ago

Please see the README. The provided checkpoints are for 10% ZINC + 10% PubChem, not the full MoLFormer-XL weights.