CrossLab-NMC v2

Overview

As scientific generation gets cheaper, the bottleneck moves to verification. Models can produce many plausible claims, yet errors remain hard to detect without checking the underlying experimental evidence. In experimental science, that evidence is usually a distribution over conditions, materials, and labs rather than a single ground-truth value, and producing new evidence can require a physical experiment.

CrossLab-NMC is a full-text, provenance-linked corpus for evaluating scientific claims against existing experimental evidence. It contains 2,005 specific-capacity measurements from 227 open-access battery papers across labs, predominantly on NMC cathodes, each linked to its source passage and to the material and experimental conditions needed to determine whether two experiments are comparable. Measurements are extracted from full text rather than abstracts, with the reported conditions kept attached to each result and mapped to a schema developed with the help of domain experts. Normalized fields make measurements comparable across papers, while raw fields preserve the source wording. The evidence supports judging whether a claim is directly observed, consistent with the record, outside the observed range, or unsupported by comparable experiments, with a traceable path back to the source passage.

Scientific claim
       ↓
Verification query
       ↓
Comparable experimental records
       ↓
Cross-paper aggregation
       ↓
Verdict and evidence
       ↓
DOI and source passage

For example, a claim that "NMC811 delivers 250 mAh/g first-discharge at 0.1C to a 4.3 V cutoff in a room-temperature, for a Li half cell with a carbonate electrolyte" can be evaluated against 13 comparable measurements from 6 papers. The paper-level median is 201.8 mAh/g and the maximum is 227.27 mAh/g, so the corpus contains no comparable observation supporting ≥250 mAh/g.

Data structure

The records subset contains 2,005 measurements across 59 fields.

Field group	Contents
Provenance	DOI, paper metadata, source license, provenance type, source passage
Results	Capacity, capacity direction, qualifier, cycle context
Materials	Composition, material family, modification, morphology
Test conditions	C-rate, voltage window, cell configuration, counter electrode, temperature
Electrolyte	Salt, concentration, solvent, additives, reported composition
Electrode	Binder, conductive additive, active-material fraction, electrode ratio, current collector, mass loading
Synthesis	Reported synthesis method

The verification_examples subset contains 12 worked claim checks, including the query, matching record IDs, evidence summary, structured answer, derived statistics, and reproducible filters.

Every measurement has paper-level provenance and a source passage. Values are not inferred from expected material behavior, and rows whose reported value could not be located in their source sentence were dropped; for values cited from a table, the passage is the table caption. No figure-sourced values are included.

Data curation

Papers were processed individually from full text so that measurements remained bound to the conditions reported with them. The extraction schema and normalization rules were developed with the help of domain experts and applied consistently across the corpus. Normalized fields support filtering and cross-paper comparison, while raw fields preserve the terminology and experimental context reported by the source.

A manual audit of 73 distinct rows found an approximately 2.5% correction rate. Each sampled row was checked against the source paper, with particular attention to the capacity value, material, C-rate, and upper voltage cutoff. This is a sample audit rather than a verification of every row.

Where cross-paper statistics are reported, measurements are reduced to the paper level so that multiple measurements from one paper do not automatically receive the same weight as independent papers. Last author is retained as a coarse provenance proxy, not as a definitive laboratory identifier.

The corpus is curated rather than exhaustive. Published experiments are subject to selection and publication bias, and many experimental conditions are incompletely reported. Cohort statistics therefore describe the observed literature record rather than the full space of possible experimental outcomes.

Expanded documentation

Three evidence concepts, kept distinct

Conflating these is the most likely way to misread an answer. Matching records establish the queried cohort, satisfying the experimental conditions, which is what makes them comparable rather than what makes them agree; evidence_record_ids lists them, and [] means the cohort was searched and returned nothing, as distinct from an empty cell meaning unknown. Supporting records are the subset that affirmatively bears on the specific claim, and evidence status is split three ways rather than two since a record that fails to support a claim is not one that contradicts it. Directly reported is stronger still, meaning the claimed value itself appears in a record rather than being inferred from where the cohort sits.

Verdict vocabulary

Verdict	Meaning
directly_reported	A matching record explicitly reports the claimed value
consistent_with_record	The claim sits within the observed evidence but is not itself reported
outside_observed_range	The claim sits outside the values observed in the matching cohort
interval_narrower_than_between_paper_scatter	The claimed interval is tighter than the spread between papers
no_comparable_evidence	No record matches the requested conditions
no_supporting_record	Matching records exist, but none affirmatively supports the claim
not_demonstrated_in_corpus	Records exist but are not a condition-matched population, so only the threshold question is answered
records_retrieved	The cohort was returned for inspection, and no claim was adjudicated

Structure and conventions

One flat table, one measurement per row, 59 columns, ordered as identity, grouping, provenance, results, materials, conditions. The seven grouping columns sit in the same order as the segments of comparison_group, so a key lines up under the columns that built it: exact material, C-rate, voltage cutoff, cycle position, cell type, temperature band, electrolyte class. Numeric columns carry their unit in the name and hold bare numbers, so voltage_bucket_V is 4.3 rather than "4.3V" and sorts and range-filters correctly, with three exceptions carrying no unit in the name (c_rate, cycle_n, year) and two that name a unit but hold a band or composite and so type as text, temp_band_C (20_28) and electrode_ratio_pct (80:10:10). Where a field has both a normalized and a verbatim form, the normalized one is for filtering and the verbatim one for citation.

Missing values follow two conventions, numeric fields left empty and text fields carrying not_reported, except comparison_group and mass_loading_is_range, which are left empty when they do not apply. comparison_group is assigned only when every facet is resolved and the row is a discharge measurement, which is 486 of the 1,838 discharge rows across 228 cohorts, since a cohort keyed on an unknown temperature is not a temperature-matched cohort; those rows keep all their data but receive no cohort key.

Extraction and audit

Extraction resolved each measurement together with its conditions in one reading rather than assembling them from separate searches, so a value and the rate, window, direction and cycle attached to it come from the same pass. The extraction schema and the cohort tolerances were curated with battery-domain guidance, and rows whose value could not be found in their own source passage were dropped rather than flagged.

The source audit ran by hand in two rounds of 40 rows, 73 distinct rows in total, reading each sampled row in full against the source paper with particular attention to the value, material, C-rate and upper cutoff. The first round drove corrections to the pipeline, and the second found 1 of 40 independently sampled rows requiring a further correction, approximately 2.5% of the audited sample. It remains a sample rather than a verification of every row in the corpus.

Remaining limitations

No quartiles are published at this number of papers, since a quartile over three to six values interpolates between two adjacent numbers and adds nothing; the paper-level medians are listed individually instead. A single cohort can contain coated, doped and unmodified material, and in one case the deliberately weaker arms of a paper's own synthesis sweep, so the modifier field records what each paper reported and the verification examples state this wherever it affects an answer. Published results carry selection and publication bias, so close agreement within a cohort does not establish a reproducibility floor. The twelve worked examples are demonstrations rather than a survey, and nine of them run on a single cohort chosen because it is the best-populated one.

Full column dictionary (59)

Identity

Column	Filled	Definition
example_id	100%	Stable citable row id, format crosslab-nmc-v2-NNNN

Grouping

Column	Filled	Definition
comparison_group	24%	Readable composite of the seven facets, DISCHARGE ROWS ONLY. reproducing a cohort from the facet columns also requires filtering capacity_type==discharge
canonical_material_id	96%	Precise stoichiometry. Use this for exact-composition cohorts, not material_family
c_rate_bucket_C	85%	Log-spaced bin on the standard C-rate series C/20 to 10C, boundaries at the geometric midpoints between adjacent standard rates. An off-series rate is folded to the nearest standard rate (0.3C reads 0.2, 3C reads 2), so a bin is a neighbourhood rather than an exact rate and the exact value is always in c_rate. Rates outside the C/20 to 10C range entirely are left unbinned. Numeric, so it sorts and range-filters.
voltage_bucket_V	78%	UPPER cutoff only, to nearest 0.05 V. The lower cutoff is NOT in the bucket and varies within it, so pair with voltage_window_lower_V when window width matters
cycle_bucket	66%	Cycle position as a numeric band: 1, 2_10, 11_50, 51_100, 101_200, 201_500, 501+, or not_reported. A value the source describes as cycled without giving a number reads not_reported here; cycle_context distinguishes it from a value with no cycle information at all.
cell_type_class	90%	Normalized cell configuration: half, full, or not_reported.
temp_band_C	62%	Reported test temperature in degrees C, as a band. Bands are closed, so 30 C and 90 C do not share a bucket. A paper stating only "room temperature" with no number is placed in 20_28 without inventing a number; operating_temperature_C is blank on those rows.
electrolyte_class	80%	Rule-based from the electrolyte fields. A gel or polymer electrolyte on a carbonate base can read as carbonate_liquid; check Methods for load-bearing use

Provenance

Column	Filled	Definition
doi	100%	Source paper DOI
title	100%	Paper title
abstract	100%	Source abstract
journal	100%	Journal
year	100%	Publication year
author	100%	Full author list
last_author	100%	Last author, used as the lab proxy
corresponding_author	100%	Corresponding author
oa_url	100%	Open-access full text or PDF
source_license	100%	Licence determination for the source paper, CC-BY-4.0 throughout this release, taken from OpenAlex licence metadata at curation rather than from the publisher rights page
provenance_type	100%	Where in the paper the value was cited from. No figure-sourced values
sentence	100%	Verbatim passage the value was taken from. For the 1,938 rows with provenance_type=sentence the value string appears in this passage, and rows failing that check were dropped. For the 67 rows with provenance_type=table_caption the passage is the caption and the value generally sits in the table it names rather than in the caption text, though 5 of those captions do contain it.

Results

Column	Filled	Definition
variable	100%	Measured quantity, constant specific_capacity
value_mAh_g	100%	The capacity value, verbatim to the paper's own precision, never rounded to uniform significant figures
capacity_type	96%	Direction of the measurement, taken from the source. not_reported means the source named a capacity without stating a direction, and none is inferred for it
capacity_qualifier	100%	Paper-stated qualifier on the number. reversible and initial are not directly comparable to a plain discharge value
cycle_context	81%	How the value binds to a cycle. Re-extracted per paper from the full text
cycle_n	66%	Cycle index where determinable

Materials

Column	Filled	Definition
material_family	100%	Coarse family label: NMC811, NMC622, NMC523, NMC111 (named standard grades), NMC-other (remaining NMC stoichiometries, most with one paper each), NCA, LRLO (Li-rich layered oxide), and other (non-NMC layered oxides plus rows whose exact composition was not reported). Use canonical_material_id for exact composition
material	100%	Material as named in the paper
modifier	50%	What the paper called this sample, verbatim. It is NOT a clean modification flag: the value none appears on 145 rows as a paper-named control, about 15% of the remaining populated values are synthesis or morphology labels (single crystal, precursor codes) rather than coatings or dopants, and it is unreliable in both directions: 266 rows read not_reported while their own passage names NCM811. Read the value, do not filter on its presence
morphology_descriptors	56%	Particle morphology where stated

Conditions: test

Column	Filled	Definition
c_rate	86%	Numeric C-rate of the reported measurement
c_rate_basis_value	29%	Reference current or capacity behind the C-rate
c_rate_basis_unit	29%	Unit of c_rate_basis_value. A unitless or mis-slotted basis is uninterpretable and is set to not_reported, along with its value
voltage_window_lower_V	75%	Lower cutoff
voltage_window_upper_V	78%	Upper cutoff
voltage_reference	35%	Voltage reference, typically Li/Li+
cell_type	90%	Cell configuration as stated
counter_electrode	92%	Counter or negative electrode as reported. Usually Li metal in half cells, but some half cells use a lithium-indium alloy, which sits about 0.62 V above Li/Li+ and is recorded as Li-In rather than folded into Li metal
operating_temperature_C	52%	Test temperature, central value recovered from ranges (20 +/- 2 -> 20)
operating_temperature_raw	63%	Verbatim temperature string

Conditions: electrolyte

Column	Filled	Definition
electrolyte	83%	Salt
electrolyte_molarity_M	58%	Salt molarity
electrolyte_solvent	67%	Solvent system
electrolyte_additive	22%	Additive, e.g. FEC or VC
electrolyte_composition_raw	87%	Verbatim recipe, provenance only. Filter on the parsed columns

Conditions: electrode

Column	Filled	Definition
binder	69%	Binder
binder_wt_pct	66%	Binder fraction
electrode_conductive_additive	74%	Conductive additive. Plurals and casing are folded (CNTs -> CNT), but vendor and abbreviation strings are only partly normalized, so carbon black (Super P, Timcal) sits apart from Super P, and acetylene black, AB sits apart from acetylene black. Read the 36 distinct values before filtering on equality. Genuine multi-component mixes are kept verbatim
conductive_additive_wt_pct	70%	Additive fraction
active_material_wt_pct	73%	Active-material fraction. Where it disagreed with the verbatim recipe in electrode_ratio_raw by more than 0.5 points it was recomputed from that recipe; a smaller disagreement keeps the paper-stated rounded figure, so a value such as 92 can sit beside a 92.5:4.5:3 recipe
electrode_ratio_pct	57%	Electrode composition on a 100 basis, so 8:1:1 and 80:10:10 land together. Active material is always first, but the value is not always three-part: 133 of the populated rows carry two to five components, because a paper reporting two additives keeps both. Split on the separator rather than assuming three fields
electrode_ratio_raw	74%	Verbatim ratio, deliberately preserving each paper's own scale and wording
electrode_solvent	55%	Slurry solvent
current_collector	64%	Current collector
mass_loading_raw	50%	Verbatim mass-loading string
mass_loading_mg_cm2_est	48%	Areal mass loading. Midpoint where the paper gave a range, flagged by mass_loading_is_range
mass_loading_is_range	50%	true where mass_loading_mg_cm2_est is the midpoint of a reported range, false where a single value was reported, blank where no mass loading was reported. 29 rows report a loading that could not be parsed: those read false with mass_loading_mg_cm2_est blank and the text kept in mass_loading_raw

Conditions: synthesis

Column	Filled	Definition
synthesis_method	47%	Synthesis route

License

The curation layer is released under CC-BY-4.0. All 227 source papers were selected as CC-BY and the determination was rechecked against OpenAlex license metadata before this release. Each measurement carries its DOI and provenance information; copyright in the underlying articles remains with their publishers.

Citation

@dataset{crosslab_nmc_2026,
  title   = {GenData-Research/crosslab-nmc},
  author  = {GenData},
  year    = {2026},
  version = {2.0},
  url     = {https://huggingface.co/datasets/GenData-Research/crosslab-nmc}
}