OpenMol/Reagent_Selection

---
license: mit
---

## Data Source
Suzuki-Miyaura Dataset. Converted from [ChemLLMBench](https://github.com/ChemFoundationModels/ChemLLMBench)\, which can be referred to Suzuki High-Throughput Experiment (HTE) dataset.\
We build the instruction dataset strictly following ChemLLMBench instruction template. But We convert the output into 'integer' (as a ranking task).\
In general, three different tasks involved:
- `reagent selection` (train: #2387, test: #100)
- `solvent selection` (train:#1155, test: #100)
- `ligand selection` (train:#413, test: #100)

## Data Stats
Train: 3955 \
Test: 300

Example:
```json
{
'task': 'reagent_selection',
'molecules': {'selfies': LIST[SELFIES]], 'smiles': LIST[SMILES]},
'messages': [
{
'content': 'You are an expert chemist. Given selected one reactant, two reagents and solvent of a Suzuki reaction, predict the optimal reactant that maximize the yield with the rest of reaction components by using your experienced chemical reactant selection knowledge. No explanations and other information. Only return the reactant smiles from the given list. Please strictly follow the format, the template are provided as follows:\nGiven the rest of reaction components:\nreactant: reactant 1\n ligand: ligand 1\n base: base 1\n solvent: solvent 1 \nReactant list: reactant1,reactant2,reactant3,.etc\nOptimal reactant: reactant2 ',
'role': 'system'
},
{
'content': 'Given the rest of reaction components:\nreactant: <molecule_2d>\nligand: <molecule_2d>\nsolvent: <molecule_2d>\nbase: <molecule_2d> \nReactants list for selection:\n<molecule_2d>,<molecule_2d>,<molecule_2d>\nOptimal reactant:\n',
'role': 'user'
},
{
'content': '2',
'role': 'assistant'
}
],
'id': 0
}
```