C-H-Formula

This repository contains NMRTrans checkpoints for C-NMR + H-NMR + Formula input.

Checkpoint	Code branch	Epoch	Top-1	Top-3	Top-10	Token acc.	Tanimoto similarity
nmrtrans-c-h-nmr-formula.ckpt	main	5409	0.4447	-	-	0.7229	0.7569
nmrtrans-c-h-nmr-formula-nmrexp.ckpt	bugfix	2719	0.4924	0.6587	0.7403	0.7726	0.8159

The updated checkpoint was trained on the merged NMRTrans + NMRExp corpus with the corrected Set Transformer peak mask and the expanded 223-token SMILES vocabulary. Its validation-best checkpoint was reached after approximately 4 days 18 hours 41 minutes on four H200 GPUs. The reported result uses the cleaned 113,323-record subset of the newly held-out test split. Because the corpus and split changed, the two rows are not a controlled single-factor ablation.

Use the original checkpoint with the repository main branch. Use the NMRExp-merged checkpoint with the bugfix branch, which includes the matching mask behavior, vocabulary, and configuration.

git switch bugfix

huggingface-cli download little1d/C-H-Formula \
  nmrtrans-c-h-nmr-formula-nmrexp.ckpt \
  --local-dir checkpoints/pretrained

python src/test.py \
  --config_path configs/config.yaml \
  --ckpt_path checkpoints/pretrained/nmrtrans-c-h-nmr-formula-nmrexp.ckpt \
  --features c_nmr,h_nmr,formula