Add: A Hugging Face pretrained molecular generator+ train_test_split implementation

[m21hm9]

This PR adds HFPretrainedMolecularGenerator: a sklearn-style fit / generate wrapper for Hugging Face pretrained molecular generators, in the same spirit as HFPretrainedMolecularEncoder. The existing generators (LSTM, MolGPT, DiGress, and the rest) are left as they are.

"API": We know some of these Hub models don’t actually speak SMILES internally, MolGen works in SELFIES, Molexar works in Fragment-SELFIES. So, we convert SMILES → SELFIES (or Fragment-SELFIES) on the way in, and convert back to SMILES on the way out.

"Supported models": NovoMolGen and GP-MoLFormer (causal LM / SMILES), MolGen-large and MolGen-large-opt (seq2seq SELFIES), Molexar-10M-base and Molexar-10M-omni (Fragment-SELFIES via the official engine). Extra deps stay in README Additional Packages, not in pyproject.toml. GP-MoLFormer needs transformers<=4.56.2; Molexar needs transformers>=5.8, in which they cannot share an environment.

"Handled": Invalid training SMILES raise with an index (no silent skip). Failed MolGen/Molexar decodes are dropped instead of padded with "", with a got k/n valid SMILES warning. MolGen fine-tuning is denoising, not identity copy. GP-MoLFormer generation is patched for transformers 4.56 cache issues, scaffold= must be an open prefix like "c1cccc", not a closed ring.

[m21hm9]

We added more README / install details: Hub links for the supported generators, extra-package notes and importantly we also changed from GP-MoLFormer to SAFE-GPT.

Two problems exist in GP-MoLFormer:
-IBM’s remote code does not play well with current transformers 5.x. It needed transformers<=4.56.2, while Molexar needs transformers>=5.8, so they could not live in the same environment.
-Even after the cache workaround, sampling quality was not good enough (especially scaffold completion)

So we replaced that scaffold / de novo causal-LM slot with SAFE-GPT(https://huggingface.co/datamol-io/safe-gpt). It is a normal GPT-2 model, so it can use standard Hugging Face generate() on transformers 5.x and sit next to NovoMolGen / MolGen / Molexar. scaffold= can be a full SMILES like "c1ccccc1"; we convert it to SAFE internally. Extra dep is just pip install safe-mol.

[m21hm9]

We added train test split on the dataset. Total 4 methods:

-random: baseline. Use as a control; scores are often a bit high.
-scaffold: Bemis–Murcko groups. The same scaffold does not appear in both splits (unseen-framework generalization). Optional use_csk=True for generic cyclic skeletons.
-butina: Taylor–Butina clusters on Morgan fingerprints (radius 2, 2048 bits). Neighbors are pairs with Tanimoto similarity ≥ 0.65 (not DeepChem’s distance cutoff). We use a sparse neighbor graph + exclusion spheres so we do not build the full distance matrix. Whole clusters go to train or holdout. Invalid SMILES raise, same as scaffold.
-size: split by heavy-atom count (small → large by default). This is only the split, not SizeShiftReg coarsening / CMD. mode="sizeshiftreg" is the paper protocol (smallest 50% train, largest 10% holdout).

Thanks!

[m21hm9]

We fixed causal LM scaffold generation and cleaned up the pretrained generator tests:

-max_new_tokens: NovoMolGen / SAFE-GPT used Hugging Face max_length as a total sequence cap (prefix + new tokens). A long SAFE/SMILES scaffold ate that budget (53-token prefix → only 11 new tokens). They now default to max_new_tokens=generate_max_length (64 new tokens, prefix does not count). max_length= is still an explicit total-length override. MolGen decoder max_length and Molexar max_new_tokens are unchanged.
-SAFE-GPT save/load: tokenizer.save_pretrained cannot serialize the custom pre-tokenizer. Local save writes tokenizer.json via SAFETokenizer.save_pretrained.
-tests: replaced pytest files with script-style pretrained_{model}.py (novomolgen, molgen, molexar, safe_gpt), same pattern as lstm.py. SAFE-GPT is tested with a short scaffold (c1ccccc1) and a long one under the default 64 new-token budget, not by raising the cap to 128.

Please let me know if there is any issues!