Dataset Access

LeMat-Synth publishes two datasets on the HuggingFace Hub. Both are gated — access is granted automatically, but you must request it once and be authenticated.

Dataset	Contents	Use it for
LeMat-Synth	Extracted synthesis procedures and figures, one row per synthesis	Using the published data directly — no extraction, no API keys, no cost
LeMat-Synth-Papers	The intermediate corpus: ~81k papers as text, one row per paper	Finding papers to run the pipeline on, or reproducing the extraction

Tip

If your question is "has anyone already extracted this?", start with LeMat-Synth and Tutorial 1. For a great many use cases that is the entire job, and it costs nothing.

Get access

1 — Request access. Open either dataset page while signed in to HuggingFace and accept the terms. Approval is automatic and immediate.

2 — Install the HuggingFace CLI.

pip install -U "huggingface_hub"    # or, on macOS: brew install hf

3 — Authenticate.

hf auth login

Alternatively, put a token from huggingface.co/settings/tokens into .env as HF_TOKEN=hf_… — the CLI, the notebooks and the Hydra deployment scripts all read it from there.

Load it

from datasets import load_dataset

synth = load_dataset("LeMaterial/LeMat-Synth", split="train")
print(synth[0]["synthesis"])

Tutorial 1 goes further — slicing the dataset by synthesis method, material category and judge score, and turning a row back into a GeneralSynthesisOntology object.

Tutorial 2 covers LeMat-Synth-Papers: filtering the corpus by category and keyword, with whole-word matching and an optional LLM relevance filter, to assemble your own input set.

Using the corpus as pipeline input

The Hydra deployment scripts read LeMat-Synth-Papers directly — it is the default data loader:

uv run examples/scripts/deployment/extract_synthesis_procedure_from_text.py
uv run ... data_loader.number_of_samples=10          # cap the run
uv run ... data_loader.architecture.split=chemrxiv   # pick a split

See Configuration & Models for the other data loaders (local text folders, the annotation subset).

Troubleshooting

Note

GatedRepoError or a 401. You are either not authenticated (hf auth login) or the access request has not been accepted on the dataset page. Both datasets are gated separately — accepting one does not grant the other.