transfer-learn

REINVENT4 Transfer Learning

Fine-tune a prior model on a SMILES dataset by writing a TOML config and running the reinvent CLI directly. No MCP server needed — activate the reinvent4 conda env first.

Workflow

1. Resolve paths

readlink -f <relative_path>

All paths in the TOML must be absolute.

2. Determine parameters

• input_model_file (required): absolute path to .prior or .model file. Use $0 if provided.
• smiles_file (required): absolute path to SMILES/CSV training data. Use $1 if provided.
• output_model_file: where to save the fine-tuned model (default: <workdir>/model.model)
• num_epochs: training epochs (default 50)
• batch_size: batch size (default 50)
• save_every_n_epochs: checkpoint frequency (default 10)
• generator: reinvent, libinvent, linkinvent, or mol2mol (default reinvent)
• standardize_smiles: canonicalise and sanitise SMILES before training (default true; set false only if pre-cleaned)
• mol2mol_pairs: for mol2mol generator — source→target SMILES pairs ({source_smi: target_smi, ...}); provide as a separate CSV with source and target columns if large
• device: cpu (default) or cuda:0

3. Validate input SMILES

Run a quick sanity check with Python/RDKit before committing to a run:

from rdkit import Chem

errors = []
with open("/absolute/path/to/smiles_file.smi") as f:
    for i, line in enumerate(f, 1):
        smi = line.split()[0].strip()
        if not smi:
            continue
        mol = Chem.MolFromSmiles(smi)
        if mol is None:
            errors.append((i, smi))

print(f"Total lines: {i}, Invalid: {len(errors)}")
for idx, smi in errors[:10]:
    print(f"  Line {idx}: {smi}")

Report any invalid entries. Offer to write a cleaned file (skip invalid lines) if needed.

4. Create the output directory

mkdir -p <workdir>

Suggested: <model_parent>/reinvent_runs/tl_<YYYYMMDD_HHMMSS>/

5. Write the TOML config

Write <workdir>/config.toml with the Write tool: