The motivation for including the original SMILES strings in the augmented dataset comes from two considerations:

1. Analogy to data augmentation in Computer Vision
   In the vision domain, it is standard practice to include both the original images and their augmented variants (rotations, flips, scaling, color jitter, etc.) when training CNN models. The original samples provide a stable reference distribution, while augmented samples improve robustness.
   By analogy, including the original SMILES ensures that the model retains exposure to the true, data distribution of ChEBI, rather than relying solely on augmented variants.

2. Chemical-domain constraints on SMILES generation
   A more domain-specific reason is that some SMILES strings present in ChEBI can be parsed by RDKit but cannot be regenerated by RDKit’s SMILES writing algorithms.
   This happens due to differences in how RDKit handles certain structural representations.
   For example, RDKit often removes implicit hydrogens or normalizes parts of the representation during canonicalization. As a result, certain SMILES in ChEBI that contain specific forms of implicit hydrogens or uncommon notations will not appear in augmented outputs generated by RDKit.
   Therefore, these original SMILES must be included in the training set, because they represent valid chemical structures found in ChEBI but would otherwise be lost during augmentation.

Below is an example illustrating such a case:

and check its generated SMILES here: https://github.com/ChEB-AI/python-chebai/wiki/SMILES-Augmentation#snippet-of-augmented-smiles

below program reinforces this theory

import random
from itertools import cycle, permutations, product

from rdkit import Chem

AUG_SMILES_VARIATIONS = 1000000000


def generate_augmented_smiles(smiles: str) -> list[str]:
    mol: Chem.Mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        return [smiles]  # if mol is None, return original SMILES

    # sanitization set to False, as it can alter the fragment representation in ways you might not want.
    # As we don’t want RDKit to "fix" fragments, only need the fragments as-is, to generate SMILES strings.
    frags = Chem.GetMolFrags(mol, asMols=True, sanitizeFrags=False)
    augmented = set()

    frag_smiles: list[set] = []
    for frag in frags:
        atom_ids = [atom.GetIdx() for atom in frag.GetAtoms()]
        random.shuffle(atom_ids)  # seed set by lightning
        atom_id_iter = cycle(atom_ids)
        frag_smiles.append(
            {
                Chem.MolToSmiles(frag, rootedAtAtom=next(atom_id_iter), doRandom=True)
                for _ in range(AUG_SMILES_VARIATIONS)
            }
        )
    if len(frags) > 1:
        for perm in permutations(frag_smiles):
            for combo in product(*perm):
                augmented.add(".".join(combo))
                if smiles in augmented:
                    print("Found original SMILES in augmented set.")
                    break
    else:
        augmented = frag_smiles[0]

    if smiles in augmented:
        print("Found original SMILES in augmented set.")
    else:
        print("Original SMILES NOT found in augmented set.")


if __name__ == "__main__":
    test_smiles = "[F-].[H][N]([H])([H])[Ag+][N]([H])([H])[H]"
    generate_augmented_smiles(test_smiles)

I would argue that the computer vision approach cannot be directly applied to SMILES. In computer vision, there is usually a "real" image and "distorted / modified" versions of the image. In SMILES, I don't see a clear distinction why the ChEBI-provided SMILES should be the "real" one that needs to be kept - all other SMILES strings are just as valid. The question then becomes: Which kind of SMILES do our users use? Is it the ChEBI-style SMILES? Is it the RDKit-canonical SMILES? I don't know.

Fun fact: The reason why we can't reproduce the ChEBI-style SMILES is that they have their own library for that: https://github.com/chembl/libRDChEBI (currently, I don't see a use case where we would need to reproduce ChEBI's exact SMILES, but if it arises, we should use this library).

Anyway, since I don't have a better suggestion, let's keep the original ChEBI-SMILES for now.