Materials science (energy storage devices) · GenData-Research
crosslab-nmc
A benchmark and evidence corpus for evaluating scientific claims about battery cathodes against literature evidence.
What crosslab-nmc is
CrossLab-NMC (released in later versions as the Scientific Verification Benchmark: NMC Cathodes) is a full-text, provenance-linked corpus developed by GenData-Research. Designed for materials science and energy storage research, it evaluates whether language models can verify scientific claims against experimental literature rather than hallucinating or overreaching beyond what original papers establish.
Experimental battery results are highly context-dependent, and primary literature often omits crucial test conditions or reports parameters inconsistently across sections. This corpus tackles the problem of evidentiary verification by pairing battery cathode claims with ground-truth literature evidence, tracking whether models adhere to reported evidence or make unsubstantiated inferences to support claims.
The resource combines a set of specific-capacity measurements curated from open-access battery papers with an evaluation benchmark assessing language model performance across scientific verification tasks.
How it was made
The dataset was compiled from 227 open-access battery papers. Specific capacity values are linked directly to source sentences or table captions, while corresponding test parameters—such as material specifications, electrolyte composition, electrode formulation, and testing conditions—are grounded in connected text passages like methods sections. Every reference citation was programmatically checked to confirm it appears verbatim in the underlying source document.
The evaluation suite tests model answers against 50 claims regarding lithium nickel manganese cobalt oxide (NMC) cathodes. Each verification record captures the model's verdict, verbatim excerpt, and relied-upon value alongside reference values, verbatim evidence quotes, and categorical outcome labels indicating whether the model remained strictly within the evidence or engaged in silent overreach.
Key facts
- Domain
- Materials science and energy storage (NMC battery cathodes)
- Task categories
- Question answering, text classification, scientific claim verification
- Source literature
- 227 open-access battery papers
- Benchmark split (dev)
- 150 scored answers across 50 claims
- Records split (full)
- 1,991 curated experimental measurements (earlier index contains 2,005 measurements)
- Modalities
- Tabular, text
- Data format
- CSV
What you can use it for
- Scientific Claim Verification
- Test and evaluate automated models on their ability to determine if a specific claim regarding NMC cathode capacity is supported, unsupported, or unevidenced by primary literature.
- Model Overreach Detection
- Assess whether scientific language models infer unstated experimental conditions or combine separate tests inappropriately when answering materials science questions.
- Battery Evidence Retrieval
- Query structured, experiment-grounded measurements linked to source sentences and test conditions across hundreds of open-access battery studies.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- All capacity values and experimental conditions are linked to verbatim source sentences or table captions verified by script.
- It distinguishes between verified reported facts and unstated inferences or combined calculations that distort evidentiary status.
- The data schema was developed in consultation with domain experts across multiple institutions.
Watch out for
- The verification set is limited to a small development set of 50 claims within a single material class (NMC cathodes).
- Reference annotations were reviewed by the dataset authors rather than through independent domain expert review.
- Most claims categorised as non-reported or calculable have 'no comparable evidence' as the true verdict, limiting the ability to measure generalised error rates.
Questions
Can I look inside crosslab-nmc without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is crosslab-nmc under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What is meant by model 'overreach' in this dataset?
Overreach occurs when a model treats evidence as supportive even though a decisive condition is unstated in the paper, or when it treats a value that can only be derived by combining separate samples or tests as directly reported.
What cathode chemistry does the dataset focus on?
The dataset focuses on lithium nickel manganese cobalt oxide (NMC) battery cathodes.
Which models were evaluated in the benchmark data?
The benchmark includes 150 evaluation runs across Claude Opus 5, GPT 5.6 Luna, and Gemini 3.1 Pro.
How to cite it
@dataset{scientific_verification_nmc_2026, title = {Scientific Verification Benchmark: NMC Cathodes}, author = {GenData}, year = {2026}, version = {3.0}, url = {https://huggingface.co/datasets/GenData-Research/scientific-verification} }
Written from crosslab-nmc's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.