MolOptIns (ShareGPT format)

Instruction-tuning data for multi-property molecule optimization, converted to the ShareGPT conversation format used by LLaMA-Factory.

Each example asks a model to edit an input molecule so that one or more molecular properties move in a requested direction, while keeping Tanimoto similarity to the input at ≥ 0.6.

Files

File	Rows	Notes
moloptins.jsonl	538,746	One JSON object per line. Recommended — streamable, works with the dataset viewer.
moloptins.json	538,746	Identical content as a single pretty-printed JSON array. Needs ~4 GB RAM to json.load().

Both files contain byte-for-byte equivalent records (verified row by row).

Schema

{
  "conversations": [
    {"from": "human", "value": "You are an expert medicinal chemist ... Input molecule: <SMILES> ... </SMILES>\nRequested modifications: increase QED ..."},
    {"from": "gpt",   "value": "<SMILES> ... </SMILES>"}
  ],
  "task": "qed",
  "source_smiles": "CNC(=O)c1ccc(OC)c(NC(=O)NC2CCC([NH+](C)C)C2)c1",
  "target_smiles": "CNC(=O)c1ccc(OC)c(NC(=O)NC2CCCC(C)C2)c1",
  "properties": {
    "qed":   {"source": 0.6221, "target": 0.7982, "change": 0.1761},
    "logp": null, "jnk3": null, "gsk3b": null, "drd2": null, "sa": null
  },
  "similarity": 0.6964
}

• conversations — single-turn ShareGPT (human → gpt). The system prompt is baked into the human turn; there is no separate system field.
• task — +-joined property names, always consistent with the non-null keys of properties.
• properties[p].change — signed improvement, not a raw delta:
  • sa: change = source - target (SA score should go down)
  • all others: change = target - source (should go up)
• similarity — Tanimoto over Morgan fingerprints (radius 2, 2048 bits).

Properties and tasks

Six properties — qed, logp, jnk3, gsk3b, drd2, sa — combined into 63 task types (1 to 6 properties at a time).

Property	Direction	Rows
qed (drug-likeness)	increase	219,337
drd2 (DRD2 inhibition)	increase	200,180
logp (lipophilicity)	increase	164,454
sa (synthetic accessibility)	decrease	70,556
gsk3b (GSK3β inhibition)	increase	56,792
jnk3 (JNK3 inhibition)	increase	26,561

Single-property tasks are the largest blocks (qed 130,987, drd2 108,921, logp 83,879); the tail runs down to 12 rows for the full 6-property combination.

Data quality

Verified over the full 538,746 rows (RDKit checks on a 1,078-row systematic sample):

• 0 unparseable SMILES (source or target)
• 0 rows where source_smiles == target_smiles
• 0 rows violating the similarity ≥ 0.6 constraint (min 0.6020, median 0.6515, mean 0.6659)
• 0 rows where task disagrees with the non-null properties keys
• 0 rows with a wrong-direction property change
• similarity reproduces exactly from RDKit Morgan/Tanimoto (max abs. diff 0.0000)

Sequence lengths are short: prompts average 730 characters (max 877), responses 64 (max 109). A cutoff_len of 512 tokens is sufficient.

Known limitations

• Duplicates. 74,939 rows (13.9%) repeat an existing (source_smiles, target_smiles, task) triple, up to 4× for the same triple. Deduplicate if that matters for your setup. (Separately, 56.4% of (source, target) pairs repeat across rows — this is expected, since one pair legitimately appears under several task labels, e.g. qed, logp and logp+qed.)
• No held-out split. This is a single undivided file. If you split it yourself, group by source_smiles: there are 122,215 unique source molecules reused 4.41× on average and up to 1,404× for the most frequent one, so a random row-level split leaks source molecules into the test set.
• Row order is grouped by task, not shuffled. Truncating the head of the file yields a single-task subset, not a representative sample — shuffle before taking any subset.
• One instruction phrasing only. Every row uses the same "seen" property wording (QED, LogP, DRD2 inhibition, ...). There are no paraphrased/"unseen" instructions, so this data does not support instruction-generalization evaluation on its own.
• 12.1% of target SMILES are not in RDKit canonical form. Canonicalize both sides before doing any string-level comparison at evaluation time.

Usage

from datasets import load_dataset

ds = load_dataset("AlexanderWang915/MolOptIns", split="train")
print(ds[0]["conversations"])

With LLaMA-Factory, add to data/dataset_info.json:

"moloptins": {
  "hf_hub_url": "AlexanderWang915/MolOptIns",
  "file_name": "moloptins.jsonl",
  "formatting": "sharegpt",
  "columns": { "messages": "conversations" }
}

Do not add a "system" column — the records have no system field and LLaMA-Factory's converter will raise a KeyError.

Provenance

Converted to ShareGPT format from the MolOptIns molecule-optimization instruction dataset. The conversion only reformats the records: it builds the prompt/response turns from source_smiles, target_smiles and task, and carries the original properties and similarity values through unchanged. No molecules, property values or pairs were added, removed or recomputed.

Note: the upstream release and license for the original MolOptIns data have not been confirmed here. Please cite the original authors and check their license terms before redistributing or using this data. If you are the upstream author and want attribution corrected or this mirror removed, open a discussion on this repo.