nfsrulesFR/mega-moledit-large

---
license: gpl-3.0
task_categories:
- text-generation
language:
- en
tags:
- chemistry
- molecular-editing
- drug-discovery
- smiles
- molecule-generation
pretty_name: MEGA Molecular Editing Dataset (Large - 62M)
size_categories:
- 10M<n<100M
dataset_info:
features:
- name: task_id
dtype: int64
- name: prompt
dtype: string
- name: input_smiles
dtype: string
- name: output_smiles
dtype: string
- name: action_type
dtype: string
- name: edit
dtype: string
- name: target_delta
dtype: float64
- name: SA_delta
dtype: float64
- name: MW_delta
dtype: float64
- name: QED_delta
dtype: float64
- name: murcko_scaffold_retained
dtype: bool
splits:
- name: train
num_bytes: 10576225637
num_examples: 28219060
- name: validation
num_bytes: 1175208625
num_examples: 3135462
- name: train_neg
num_bytes: 8843010060
num_examples: 23684301
- name: validation_neg
num_bytes: 982619844
num_examples: 2631604
download_size: 6506509462
dataset_size: 21577064166
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: train_neg
path: data/train_neg-*
- split: validation_neg
path: data/validation_neg-*
---

# MEGA: A Large-Scale Molecular Editing Dataset for Guided-Action Optimization

Large-scale annotated molecular editing dataset with 57M examplesfor training models to modify molecular structures based on natural language instructions.

**Paper**: [MEGA: A Large-Scale Molecular Editing Dataset for Guided-Action Optimization](https://openreview.net/pdf?id=MaS7e2EVFm)
**Official Repository**: [https://github.com/nfsrules/MEGA-moledit](https://github.com/nfsrules/MEGA-moledit)

## Dataset Structure

Each example will contain:
- `task_id`: Task identifier
- `prompt`: Natural language instruction
- `input_smiles`: Input molecule
- `output_smiles`: Target molecule
- `action_type`: Edit operation type
- `edit`: Specific edit applied
- `target_delta`: Change in target property
- `SA_delta`: Change in Synthetic Accessibility
- `MW_delta`: Change in Molecular Weight
- `QED_delta`: Change in Drug-likeness
- `murcko_scaffold_retained`: Scaffold preservation flag

## Supported Tasks

| Task ID | Description |
|---------|-------------|
| 101 | Increase water solubility |
| 102 | Decrease water solubility |
| 103 | Increase drug-likeness |
| 104 | Decrease drug-likeness |
| 105 | Increase permeability |
| 106 | Decrease permeability |
| 107 | Increase hydrogen bond acceptors |
| 108 | Increase hydrogen bond donors |
| 201 | Increase solubility + HBA |
| 202 | Decrease solubility + increase HBA |
| 203 | Increase solubility + HBD |
| 204 | Decrease solubility + increase HBD |
| 205 | Increase solubility + permeability |
| 206 | Increase solubility + decrease permeability |

**Splits**: `train` (31M positive + 26M negative), `validation`

## Smaller Variant

Check out the smaller version: [nfsrulesFR/mega-moledit-522K](https://huggingface.co/datasets/nfsrulesFR/mega-moledit-522K)

## Trained Models

Llama 3 8B-based models for molecular optimization:

- **MEGA-SFT**: [nfsrulesFR/mega-sft](https://huggingface.co/nfsrulesFR/mega-sft) - Supervised fine-tuning model
- **MEGA-GRPO**: [nfsrulesFR/mega-grpo](https://huggingface.co/nfsrulesFR/mega-grpo) - Tanimoto-GRPO optimized model

## Citation

```bibtex
@article{
fernandezillouz2025mega,
title={MEGA: A Large-Scale Molecular Editing Dataset for Guided-Action Optimization},
author={Nelson Fernandez and Maxime Illouz and Luis Pinto and Entao Yang and Habiboulaye Amadou Boubacar},
journal={Under review at International Conference on Learning Representations},
year={2025},
url={https://openreview.net/pdf?id=MaS7e2EVFm}
}
```