summykai/chemistry-sft-ultra

---
language: en
license: other
license_name: various
license_details: This dataset combines multiple source datasets with varying licenses. Users should review the original source datasets and their respective licenses.
tags:
  - chemistry
  - science
  - smi
  - molecules
  - instruction-tuning
  - sft
  - question-answering
  - text-generation
annotations_creators:
  - expert-generated
  - machine-generated
language_creators:
  - found
language_details: en-US
pretty_name: Chemistry SFT Ultra
size_categories:
  - 1M<n<10M
source_datasets:
  - smolininstruct
  - omnichem
  - moleculeqa
  - chemrxiv
  - tulu-3-sft-mixture
  - numinamath-cot
  - textbookreasoning
task_categories:
  - question-answering
  - text-generation
task_ids:
  - extractive-qa
  - open-domain-qa
  - text2text-generation
configs:
  - config_name: default
    data_files:
      - split: train
        path: chemistry_mix.parquet
    dataset_info:
      features:
        - name: messages
          list:
            - name: role
              dtype: string
            - name: content
              dtype: string
        - name: metadata
          struct:
            - name: source
              dtype: string
            - name: split
              dtype: string
            - name: task
              dtype: string
            - name: language
              dtype: string
            - name: licence
              dtype: string
            - name: chemistry
              dtype: bool
    splits:
      - name: train
        num_examples: 1370322
        num_bytes: 0
    download_size: 0
    dataset_size: 0
---

# Chemistry SFT Ultra

Modern chemistry fine-tuning data built from seven rigorously curated open datasets, merged into a single English-only corpus with reproducible processing and analysis.

## Dataset Summary

This repository curates seven open-source instruction and QA datasets into a single, cleaned, deduplicated, English-only training corpus. Each source is processed by purpose-built scripts, validated structurally, and profiled with quantitative heuristics and chemistry-specific quality checks (RDKit).

The final corpus contains **1,370,322 rows** with:
- Deterministic processing and validation
- **Very low duplication** (≈0.00% estimated)
- Strong **chemistry signal** (SMILES detection) and **RDKit validity** (≈99.72% parse/canon on a 200k sample after light salvage)
- Balanced **source/task composition** (entropy metrics)

## Source Datasets

| Dataset | Target Rows | Actual Rows |
| --- | ---: | ---: |
| SMolInstruct | 440 000 | 440 000 |
| OmniChem | 199 000 | 199 000 |
| MoleculeQA | 60 000 | 60 000 |
| ChemRxiv-Train | 137 635 | 137 635 |
| Tulu-3 SFT Mixture | 370 000 | 370 000 |
| NuminaMath-CoT | 170 000 | 170 000 |
| TextbookReasoning | 97 000 | 97 000 |
| **Merged Chemistry Mix** | — | **1,370,322** |

Row totals reflect post-filtering counts and final merge deduplication + language filtering (English probability ≥ 0.98).

## Dataset Structure

Each example contains:
- `messages`: A list of message dictionaries with `role` and `content` keys
- `metadata`: A dictionary with source information including:
  - `source`: Original dataset name
  - `split`: Dataset split (typically "train")
  - `task`: Task type (e.g., "retrosynthesis", "property_prediction", etc.)
  - `language`: Language code (en)
  - `licence`: License information
  - `chemistry`: Boolean flag indicating chemistry content

### Data Fields

```python
{
    "messages": [
        {"role": "user", "content": "..."},
        {"role": "assistant", "content": "..."}
    ],
    "metadata": {
        "source": "SMolInstruct",
        "split": "train",
        "task": "retrosynthesis",
        "language": "en",
        "licence": "MIT",
        "chemistry": True
    }
}

Dataset Analysis

Latest Metrics

chemistry_mix:
  Rows: 1,370,322
  Turns -> min 2, max 294, avg 2.06, >2 turns 0.75%
  Empty rows: 0 | SMILES-detected rows: 507,712
  LexDiv: 0.8066 (unique toks: 2,270,292)
  Exact dup pairs: 0 | unique pairs: 1,370,322
  Source entropy: 2.561 | Task entropy: 3.354
  RDKit: parse=99.72%, canon=100.00%, scaffold_div=0.44, |charge|≥3=0.00% (salvage 199431/200000)
  Near-dup est.: 0.00% on 300,000 rows

Interpretation: The final corpus is largely QA (few multi-turn threads), heavy on real chemistry content (≈37% rows with detected SMILES), very low duplication, and high RDKit validity after light salvage (replacing salts/whitespace variants).

Task Distribution

The dataset contains a extensive set of chemistry-related tasks including:

• Retrosynthesis and forward synthesis
• Property prediction (ESOL, Lipophilicity, HIV, etc.)
• Molecule generation and captioning
• Name conversion (SMILES ↔ IUPAC ↔ Formula)
• Chemistry QA and multiple choice questions
• Math chain-of-thought reasoning
• Instruction following
• Retrieval QA

Processing Pipeline

All datasets were processed through a rigorous pipeline:

1. Normalization: Converting each source to a canonical format
2. Quality filtering: Removing empty or invalid examples
3. Language filtering: Ensuring English-only content (probability ≥ 0.98)
4. Deduplication: Removing exact duplicates based on (user, assistant) hash
5. Validation: Automated consistency checks
6. Chemistry validation: RDKit-based SMILES validation and salvage

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/chemistry-sft-ultra", split="train")

# Access examples
example = dataset[0]
messages = example["messages"]
metadata = example["metadata"]

# Filter for chemistry examples
chemistry_examples = dataset.filter(lambda x: x["metadata"]["chemistry"])

# Filter by task
retrosynthesis_examples = dataset.filter(lambda x: x["metadata"]["task"] == "retrosynthesis")

Considerations for Use

• The dataset is primarily English-language chemistry content
• Approximately 37% of examples contain SMILES strings
• Very low duplication rate (≈0.00%)
• High RDKit validity (≈99.72% parse success rate)
• Balanced source and task distribution

License

This dataset combines multiple source datasets with varying licenses. Users should review the original source datasets and their respective licenses. The processing and curation work is released under the same terms as the source materials.

Citation

If you use this dataset in your research, please cite:

@dataset{chemistry_sft_ultra,
  title={Chemistry SFT Ultra: A Curated Chemistry Instruction Tuning Dataset},
  author={Your Name},
  year={2024},
  url={https://huggingface.co/datasets/your-username/chemistry-sft-ultra}
}