SchemaLabs
  • Product ›
    Models
    Schema-2 Latest Model registry
    Deploy
    Platform API & Docs Soon
  • Research
  • Solutions
  • Pricing
  • Company ›
    About Contact Trust Legal
Sign in →
Legal · Model Card

Schema-2

SchemaLabs, Inc.

Last updated
August 7, 2026
Status
Latest model
Documentation of record
This Model Card
Contents
  1. 1Schema-2
  2. 1.1Model details
  3. 1.2Intended use
  4. 1.3Training data
  5. 1.4Evaluation
  6. 1.4.1Data understanding
  7. 1.4.2Cross-table relating
  8. 1.4.3Scoring and anomaly detection
  9. 1.4.4Missing data
  10. 1.4.5Continual learning
  11. 1.5Limitations
  12. 1.6Bias, risks, and fairness
  13. 1.7Regulatory context
  14. 1.8Contact, citation, and resources

Schema-2 is the latest Schema Model, a Data Language Model (DLM) for tabular data. It supersedes Schema-1. All Schema Models are listed in the Model Registry.

This page documents Schema-2: its intended use, training data, evaluation, and known limitations. It is published in support of customers' compliance obligations as deployers under the EU AI Act and analogous frameworks, and is the documentation of record for the model.

Capitalised terms used on this page (including "Schema Models," "Base Model," "Fine-Tuned Checkpoint," and "Customer Data") have the meanings set forth in the Schema Model License and the Data Processing Agreement.

1. Schema-2

1.1 Model details

NameSchema-2
ReleasedAugust 2026
Model classData Language Model (DLM)
ModalityStructured data
Parameters~2.18 billion
DistributionHosted via API, platform, and supported third-party registries; weights not distributed
LicenseProprietary
Service statusLatest; supersedes Schema-1

1.2 Intended use

Operating condition. Schema-2 is designed to run on tabular data as it is: raw tables from one or many sources, with no shared keys and no schema mapping between them, no separate labeling or annotation step, and no preprocessing before inference. Column names may be absent, cryptic, or misleading, and cells may be missing. This is part of the intended purpose rather than a tolerance: the evaluations in section 1.4 are measured under these conditions.

Primary uses.

  • Dataset understanding: one-pass data understanding over tables and databases.
  • Integration as a foundation layer for vertical and agentic AI built on structured data, including invocation at runtime.
  • Relating sources: column alignment, entity matching, and unified schema proposals across tables that share no keys.
  • Working on incomplete data: missing-value imputation for numeric and categorical cells.
  • Prediction and anomaly detection: classification and regression through model endpoints over pinned data, and anomaly scoring as an explicitly requested task. Prediction uses a target column already present in the data, selected automatically or specified by the customer. Each prediction endpoint ships a held-out evaluation report, so deployers verify performance on their own data rather than relying on the figures in this card.
  • Sequential adoption of new tasks: an endpoint takes on additional data over time without a full retrain for each one.

Use cases and industry applications are cataloged in Solutions.

Runtime and agentic AI. Schema-2 is intended to be callable by an automated system at inference time, with no human preparing the data first. Removing the human preparation step does not remove the deployer's oversight duty: where an output feeds a decision affecting a person, human review remains the deployer's responsibility and is not performed by the model. See sections 1.5 and 1.7.

Users. Two groups use Schema-2 directly. Technical teams in data, ML, and platform engineering integrate it into their products and pipelines. Business teams whose work runs on tables, in finance, operations, claims, and comparable functions, use it on their data without a data-engineering step in between. Outputs also reach people with no data expertise through embedded vertical and agentic applications built on Schema-2. Deployers building for any of these audiences carry the oversight obligations in section 1.6 on behalf of users who cannot evaluate the model themselves.

Out of scope. Schema-2 is not a data-quality or data-cleansing product: imputation estimates values for cells that are absent, and anomaly scoring flags rows that do not fit the distribution, but neither corrects a value that is recorded and wrong. Also, time-ordered forecasting (time series); generative natural language tasks; image, audio, or video; safety-critical real-time decisions without human review; high-risk EU AI Act applications without human oversight; and any use prohibited by our Use Policy are out of scope.

1.3 Training data

Schema-2 was trained on a mixture of synthetic tabular datasets and real-world tabular datasets. Exact composition and dataset counts are proprietary and not disclosed.

No Customer Data was used to train Schema-2. No personally identifiable information was collected for training.

Benchmark contamination controls: evaluation suites on this page were audited against the training corpus using an identifier blocklist and a distributional near-duplicate audit.

1.4 Evaluation

Schema-2 is evaluated across ten benchmark families, grouped below into five capability areas. Evaluation follows three standing rules:

  • Competitors are never re-run. Every competitor figure is a third-party published value cited to its primary source; live leaderboard citations pin their snapshot date.
  • Published means verified. Schema-2 figures were measured on SchemaLabs' harness, under each benchmark's published protocol, on one sealed configuration per benchmark with no per-dataset tuning.
  • Protocols are never blended. Results measured under different datasets, metrics, or masking schemes are reported separately and never combined into a single figure.
Benchmark Task Protocol source
Blind sector identification Sector classification from raw cell values, no metadata No published competitor exists
Metadata independence Prediction under column-name removal Protocol of the Schema-1 evaluation (arXiv 2605.06290)
Multi-table entity matching Keyless entity matching, F1 LLM4MEM, arXiv 2604.21238
Schema matching Column correspondence, F1 LLMATCH, arXiv 2507.10897 (Valentine pairs)
ScoringBench Distributional regression, proper scoring (CRPS, CRLS) arXiv 2603.29928; output board, August 5, 2026
ADBench Zero-shot anomaly detection arXiv 2206.09426; OUTFORMER protocol, arXiv 2602.03018
Missing-data robustness Prediction under injected missingness MIRRAMS, arXiv 2507.08280
Numeric imputation Masked-cell reconstruction, NRMSE arXiv 2603.22332
Categorical imputation Masked-cell classification, macro-F1 No published competitor exists
Continual learning Retention across 50 sequential fine-tunes No published competitor exists

1.4.1 Data understanding

Blind sector identification

Methodology. 504 real-world datasets unseen in training, column names zeroed, no labels, no metadata. Schema-2 names the industry sector of each table from raw cell values alone, scored top-1 / top-5 across sectors against the uniform random baseline; accuracy is also reported at the coarser parent level, where a sibling-sector miss counts as a near-miss. Ground-truth labels come from independent judges, with ambiguous tables dropped at labeling time; the model's own prediction is never used as its own label. No prior tabular model performs this task.

Blind sector identification Top-1 and top-5 accuracy across sectors, column names zeroed, no labels, no metadata. Uniform random baseline: 0.01% top-1, 0.05% top-5.
Top-1 Top-5 Schema-2 · sector 504 sets, row of record 86.3% 91.1% Schema-2 · parent 504 sets, sibling = near-miss 86.7% 91.4% Schema-2 · two-judge set 290 sets, secondary cut 88.5% 92.3% Uniform random analytic baseline 0.01% 0.05% 0% 25% 50% 75% 100%

What the numbers mean. Point Schema-2 at a table it has never seen, headers stripped, no documentation, and it tells you what business the data comes from: right on the first guess for 6 of every 7 tables, against a 0.01% chance. That capability is automatic data cataloging: inventorying a data lake, routing incoming files to the right vertical pipeline, or triaging third-party data before a person ever opens it.

Metadata independence

Methodology. 20 OpenML numerical classification datasets under three naming conditions: full names, random-string names, names removed. Mean ROC-AUC per condition; the metric of record is the no-names value and the drop from full names (epsilon).

Prediction under column-name removal Mean ROC-AUC per naming condition.
0.96 0.94 0.92 0.90 0.88 0.86 0.84 Full names Random strings No names 0.9367 0.9289 0.9224 0.9237 0.9230 0.8658 0.8541 Schema-2 TabuLa-8B ConTextTab

What the numbers mean. Real enterprise tables carry cryptic codes, decades-old field names, or headers stripped for privacy; models that rely on reading column names lose roughly seven ROC-AUC points when that happens. Schema-2 is the only model shown whose accuracy does not fall when names are removed or replaced with noise: epsilons −0.0006 (no names, retention 100%) and −0.0013 (random strings), against 0.0709 for TabuLa-8B (92.4%) and 0.0748 for ConTextTab (91.9%). Predictions hold whether or not the data documentation is clean: no renaming, mapping, or data-dictionary work is needed before the model is useful.

1.4.2 Cross-table relating

Schema-2 relates tables that share no keys: column alignment, entity matching, and unified schema proposals. Results are reported on recognized external benchmarks.

Multi-table entity matching

Methodology. The six multi-table entity-matching datasets of LLM4MEM (arXiv 2604.21238), the current published state of the art (8B-parameter LLM backbones); metric F1. Schema-2 runs keyless: keys and row identifiers are withheld, match scores are calibrated, and the mutual-best one-to-one assignment rule is fixed before scoring.

Multi-table entity matching, F1 per dataset Six LLM4MEM datasets, F1 on a 0 to 100 scale shared across panels, higher is better. Schema-2 runs keyless; competitor values as published.
Schema-2 (keyless) LLM4MEM (Qwen2.5-7B) LLM4MEM (LLaMA3.1-8B) LLM4MEM (Falcon3-8B) MultiEM Geo 97.17 97.1 96.4 95.2 90.9 Music-20K 93.80 92.8 91.1 90.6 88.6 Music-200K 92.31 86.2 85.5 83.8 82.2 Music-2M 79.83 73.3 70.6 71.4 68.7 Person 79.13 41.8 39.3 38.0 36.5 Shopee 40.66 31.0 32.6 29.1 26.2

What the numbers mean. Merging customer, product, or supplier records across systems normally depends on shared IDs that different systems rarely share. Schema-2 finds which rows describe the same real-world entity with no keys at all, and beats the published state of the art on all six benchmarks. The margin is widest exactly where text matching gives out (Person, a 1.9x lead): records that describe the same entity in different words. For a business this is deduplication and record linkage across silos that were never designed to join.

Schema matching

Methodology. The four Valentine fabricated Musicians pairs under the LLMATCH protocol (arXiv 2507.10897, Table 2): metric F1, foreign-key matches counted as their primary keys, the Valentine harness scoring Schema-2's alignments. Every published competitor is name-aware; Schema-2 runs name-blind, a strictly harder condition.

Schema matching on the four Valentine scenarios, F1 LLMATCH protocol, higher is better, one shared 0 to 1 scale. Schema-2 is the only name-blind entry; every other method sees schema names.
Joinable (mjs-mjt) Schema-2 (blind) 0.99 LLMatch (4o-mini) 0.92 LLMatch (GPT-3.5) 0.92 ReMatch (4o-mini) 0.92 ReMatch (GPT-3.5) 0.86 Unicorn 0.92 COMA 0.91 SF 0.71 Cupid 0.33 Sem.-Joinable (msjs-msjt) Schema-2 (blind) 0.99 LLMatch (4o-mini) 0.94 LLMatch (GPT-3.5) 0.82 ReMatch (4o-mini) 0.84 ReMatch (GPT-3.5) 0.70 Unicorn 0.80 COMA 0.71 SF 0.70 Cupid 0.80 Unionable (mus-mut) Schema-2 (blind) 1.00 LLMatch (4o-mini) 0.97 LLMatch (GPT-3.5) 0.95 ReMatch (4o-mini) 0.81 ReMatch (GPT-3.5) 0.88 Unicorn 0.87 COMA 0.57 SF 0.89 Cupid 0.86 View-Unionable (mvs-mvt) Schema-2 (blind) 0.94 LLMatch (4o-mini) 0.83 LLMatch (GPT-3.5) 0.83 ReMatch (4o-mini) 0.71 ReMatch (GPT-3.5) 0.67 Unicorn 0.80 COMA 0.91 SF 0.59 Cupid 0.33

All competitor entries are name-aware; SF is Similarity Flooding. Unicorn carries a leakage caveat: trained on Valentine data, per Magneto (arXiv 2412.08194).

What the numbers mean. Before any merge, migration, or warehouse consolidation, someone has to work out which columns in one table correspond to which in another. Schema-2 produces that mapping from the data alone, without reading a single column name, and still clears every published matcher that does read them, on all four scenarios. Aggregate over the full 110-problem Valentine test set: F1 0.9870 (precision 0.9919, recall 0.9822); with column names available the same matcher reaches 1.0000. That covers data migrations and integrations where naming conventions do not line up, or where headers cannot be trusted.

1.4.3 Scoring and anomaly detection

Distributional calibration, ScoringBench

Methodology. ScoringBench (arXiv 2603.29928) evaluates distributional regression under proper scoring rules; the current output board carries 102 regression datasets, the basis reported here. Schema-2 ran locally on the official ScoringBench runner: one sealed configuration with split-conformal calibration, 5 folds on all 102 datasets, matching the board's fold protocol. Schema-2's per-dataset outputs are merged with the official published outputs of the 50 board models based on the August 5, 2026 leaderboard and every statistic is recomputed over the 51-model pool with the board's own ranking procedure.

ScoringBench CRPS, top of the merged 51-model pool* Median CRPS across 102 datasets, 5-fold, lower is better. Right column: mean rank across datasets (the board's ranking statistic).
mean rank Schema-2 0.048 1.51 nori_30m 0.2120 3.42 nori (base) 0.2453 4.18 TabPFN-3 0.3792 11.02 RealTabPFN-2.5 ft 0.3857 12.77 TabICLv2 ft 0.4018 13.76 0.0 0.1 0.2 0.3 0.4
ScoringBench CRLS, top of the merged 51-model pool* Median CRLS across 102 datasets, 5-fold, lower is better, ordered by mean rank (which is why rank order and median order disagree). Right column: mean rank across datasets.
mean rank Schema-2 0.4962 3.73 forest_diffusion_flow 0.9451 5.91 TabICLv2 ft 1.2982 10.63 CatBoost quantile 1.1556 10.85 TabICLv2 1.3288 12.79 TabPFN-3 1.2654 15.36 nori_30m 1.2707 15.45 TabPFN-2.5 ft 1.2530 16.24 0.0 0.5 1.0

* Schema-2 is not listed on scoringbench.com: the board is entered by adding a runnable model wrapper to the public ScoringBench repository, and Schema-2's weights are not distributed. The placements here are computed from the official per-dataset outputs plus Schema-2's, with the board's own ranking procedure (mean rank across datasets, Friedman and Nemenyi tests), and are reproducible from those artifacts.

What the numbers mean. Distributional regression is forecasting with honest uncertainty: not just "this claim will cost $1,200" but the full range of what it could plausibly cost. Proper scoring rules reward getting that whole range right, and on both computed rules Schema-2 places first of 51 models: forecasts for pricing, demand, and risk built on it score closer to reality than those from any model on the board.

Anomaly detection, ADBench

Methodology. ADBench (arXiv 2206.09426, NeurIPS 2022) is the standard tabular anomaly suite: 57 datasets. Schema-2's anomaly head is evaluated zero-shot on all 57 datasets under the OUTFORMER protocol (arXiv 2602.03018), three seeds; the score fuses twelve channels by weighted rank averaging under a single global rule, with no per-dataset selection. Competitor values are the per-dataset AUROC results the OUTFORMER paper publishes (Table 35, v2) over the same 57 datasets; rank and ELO are recomputed over the thirteen models shown so every cell shares one basis, with ELO averaged over 100 shuffled orderings.

ADBench zero-shot anomaly detection, 57 datasets Mean AUROC (bars, higher is better) with mean rank (lower is better) and ELO per model. All thirteen models recomputed on one basis; ordered by mean rank.
Schema-2 (DLM) zero-shot FM classical / diffusion classical deep mean rank ELO Schema-2 0.8404 4.46 1171 OutFormer 0.8242 4.65 1165 DTE-NP 0.8149 5.12 1051 TabPFN-OD 0.8101 5.44 1043 kNN 0.8086 5.92 1029 ICL 0.7936 6.29 1005 DTE-C 0.8037 6.77 1003 LOF 0.7854 6.84 987 FoMo-0D 0.7989 6.87 981 DDPM 0.7792 7.92 921 IForest 0.7575 9.25 856 GOAD 0.6882 10.31 812 DeepSVDD 0.6789 10.67 798 0.0 0.2 0.4 0.6 0.8

What the numbers mean. Anomaly scoring flags the rows that do not belong: fraudulent transactions, faulty sensor readings, data-entry errors. Zero-shot means no labeled examples of "bad" are needed; Schema-2 scores a table it has never seen and places first on every measure across the 57-dataset suite, edging a field whose published leaders are statistically tied. Anomaly scoring runs only as an explicitly requested task, never by inference.

1.4.4 Missing data

Schema-2 treats missing values as signal, not errors: the missing-value-structure pathway encodes the pattern of absence itself. Three protocols, reported separately.

Robustness under missingness

Methodology. The MIRRAMS ten-dataset suite and protocol (arXiv 2507.08280): MCAR / MAR / MNAR missingness injected independently into training and test data over a two-axis rate grid, ROC-AUC, three repeats, 750 of 750 cells. Competitor means are computed from the paper's per-dataset appendix tables.

Missing-data robustness, MIRRAMS grid Mean ROC-AUC across the ten-dataset mechanism-by-rate grid, higher is better. Competitor means computed from the paper's published per-dataset tables.
Schema-2 0.8943 MIRRAMS 0.8878 CatBoost 0.8697 SAINT 0.8608 XGBoost 0.8582 Random Forest 0.8379 TabTransformer 0.8259 SwitchTab 0.8211 NN-0 0.7681 LR-0 0.6688 0.00 0.25 0.50 0.75 1.00

What the numbers mean. This grid simulates what production data actually does: training and serving data going incomplete in different ways, at different rates, at the same time. Schema-2 predicts through it natively, with no imputation pipeline to build or maintain, and holds the top mean across the grid.

Numeric imputation

Methodology. The 20 real-world datasets and masking harness of arXiv 2603.22332: masked-cell reconstruction, NRMSE, MCAR / MAR / MNAR at 5 / 10 / 20% missing; pretraining did not see these datasets. Competitor values are the source's published Table 5 cells; each method's mean is computed from its nine published cells.

Numeric imputation: mean NRMSE 20 real-world datasets, mean over the nine MCAR/MAR/MNAR conditions, lower is better.
Schema-2 Schema-1 (DLM) LLM classical / deep tabular FM Schema-2 0.133 Schema-1 0.163 Gemini 3.0 F. 0.235 Claude 4.5 S. 0.237 Devstral 2 0.287 MiMo-V2 0.292 GPT-4.1 Nano 0.296 missForest 0.302 MICE 0.306 kNN 0.327 SAEI 0.372 SoftImpute 0.425 TabPFN 0.448 0.0 0.1 0.2 0.3 0.4 0.5

What the numbers mean. Every gap filled in a table becomes an input to the models, reports, and decisions built on it, so reconstruction error compounds downstream. Schema-2's fills carry 1.8x less error than the best published frontier LLM and 2.3x less than the best classical method, on all nine mechanism-by-rate conditions and on the mean: cleaner repaired data without choosing or tuning an imputation method per column.

No-prior split: NRMSE by mechanism The source's 9 synthetic datasets designed without prior LLM exposure, mean NRMSE, lower is better. Bold value = best per mechanism.
MAR Schema-2 0.1983 MICE 0.160 missForest 0.224 Gemini 3.0 F. 0.239 Claude 4.5 S. 0.257 MNAR 0.1920 0.429 0.425 0.422 0.433 MCAR 0.1922 0.169 0.188 0.208 0.225
Categorical imputation

Methodology. Eight of the imputation datasets with categorical cells masked; masked cells are scored as classification, macro-F1. Numeric and categorical results are never blended into one imputation number. No published competitor protocol exists in this metric, so the reference is the analytic majority-class floor.

Categorical imputation: macro-F1 vs the majority-class floor Masked categorical cells scored as classification, per dataset, higher is better. The floor is analytic (modal share).
Schema-2 Majority-class floor (analytic) german-credit 0.5076 0.3489 cleveland 0.5410 0.3702 stroke 0.8110 0.4623 hepatitis 0.9172 0.7232 student_math 0.4376 0.2939 student_port 0.4345 0.2923 chronic 0.3389 0.1023 cervical 0.3547 0.3258 Mean 0.5428 0.3649 0.00 0.25 0.50 0.75 1.00

What the numbers mean. Categorical gaps, a missing diagnosis code or an empty churn flag, cannot be averaged in; the default fix is to fill with the most common value, which systematically erases the rare categories that carry the signal. Macro-F1 punishes exactly that, and Schema-2 beats the floor on 8 of 8 datasets (mean margin +0.178), with the widest gains on the imbalanced cases (stroke +0.349, chronic +0.237) where the minority class is the one that matters.

1.4.5 Continual learning

Methodology. 50 sequential fine-tunes on distinct domains (CC18 plus TabReD used as domain tasks). Retention is a task's performance after all later cycles divided by its performance immediately after that task (computed on NRMSE), reported per task and as the mean. Gradient-boosted stacks retain 0% by construction (full retrain per task) and frozen foundation models have no update mechanism; those rows are analytic.

Retention after 50 sequential fine-tunes Task-1 and mean retention per configuration, higher is better.
100% 75% 50% 25% 0% 82.4% 80.3% 57.3% 51.2% 0% 0% task 1 mean task 1 mean task 1 mean Schema-2 full configuration Schema-2 sequential baseline GBDTs full retrain, by construction

What the numbers mean. A deployed model has to pick up new domains without forgetting the ones it already serves. After 50 sequential fine-tunes, Schema-2 keeps 80.3% of per-task performance on average, where gradient-boosted stacks keep nothing (full retrain every time) and frozen foundation models cannot learn at all: one model can follow a business across new products, regions, and datasets instead of being rebuilt for each.

1.5 Limitations

  • Probabilistic outputs: all Schema-2 outputs are probabilistic and include confidence scores. They should not be treated as ground truth.
  • Time series out of scope: Schema-2 does not perform time-ordered forecasting.
  • No automatic human review: customers are responsible for implementing human oversight where required.
  • Hosted-only: dedicated regional deployments are available as a paid option for enterprise customers; otherwise Schema-2 is available exclusively through the SchemaLabs API, platform, and supported third-party registries. Weights are not distributed.

1.6 Bias, risks, and fairness

Customers deploying Schema-2 in contexts affecting individuals (employment, lending, insurance, healthcare, education, criminal justice) bear responsibility for:

  • Auditing their data for protected-class proxies and historical bias
  • Testing Schema-2 outputs for disparate impact across protected classes
  • Implementing human oversight for high-stakes decisions
  • Complying with applicable anti-discrimination laws

Use of Schema-2 for illegal discrimination is prohibited under our Use Policy §1.1.

Schema-2 has not been formally evaluated for adversarial robustness against membership inference, model extraction, or adversarial input attacks. Our Use Policy prohibits these attack types. Customers in adversarial-environment deployments should not assume Schema-2 is hardened against such attacks.

1.7 Regulatory context

EU AI Act

Schema-2 is a general-purpose Data Language Model for tabular data and is not inherently classified as high-risk. Customer deployments may fall within high-risk categories under Annex III (credit scoring, employment decisions, insurance, healthcare diagnostics, access to essential services). Customers deploying in these contexts bear deployer-level obligations, which are in effect under the high-risk-system phase-in as of August 2026.

SchemaLabs is the Provider of Schema-2. This Model Card is the provider-level technical documentation we maintain for the model; its last-updated date (header above) and the card version tracked in the Model Registry are the audit trail of record.

For deployer obligations, see our Use Policy §2.

1.8 Contact, citation, and resources

  • Technical questions: [email protected]
  • Compliance and customer documentation: [email protected]
  • Security: [email protected]
  • General: [email protected]

Citation

SchemaLabs, Inc. (2026). Schema-2 Model Card, August 7, 2026. schemalabs.ai/model-card/schema-2

This Model Card is published as a transparency commitment to all users of the SchemaLabs Service and in support of customers' regulatory compliance obligations. It is not a contract.

SchemaLabs
Product
  • Model Registry
  • Pricing
Solutions
  • Data understanding
  • Agentic & vertical AI
  • Key-less mapping
  • All solutions
Legal
  • Privacy
  • Terms
  • Use Policy
  • Trust Center
  • All legal
Company
  • About
  • Contact
  • News
© 2026 SchemaLabs. All rights reserved. Cookie settings Foundation model for tabular data.

Cookies

We use a small, restricted set of cookies. We do not use advertising or marketing cookies, and we do not track you across third-party websites. See our Cookie Policy for details.

Cookie settings

Choose which cookies to allow. Strictly necessary cookies cannot be disabled.

  • Strictly necessary

    Essential for the website and Web App to function. Always on.

  • Remember your preferences and settings.

  • Help us understand how visitors use the site (aggregated, anonymized).