---
license: other
license_name: yhat-labs-api-terms
license_link: https://yhatlabs.com/terms
pipeline_tag: tabular-classification
tags:
  - tabular
  - tabular-classification
  - tabular-regression
  - foundation-model
  - in-context-learning
  - probabilistic-prediction
model-index:
  - name: ChakraTab
    results:
      - task:
          type: tabular-classification
        dataset:
          name: TabArena-Lite
          type: tabarena
        metrics:
          - type: Elo
            name: Elo (TabArena-Lite, full configuration, 51 datasets)
            value: 1834
          - type: Elo
            name: Elo (TabArena-Lite, medium configuration, 51 datasets)
            value: 1745
        source:
          name: TabArena leaderboard
          url: https://huggingface.co/spaces/TabArena/leaderboard
---

# ChakraTab

**ChakraTab is a tabular prediction model from [YHat Labs](https://yhatlabs.com), served as a hosted API. This repository holds the model card; there are no weights to download.**

Send a table with a target column and the rows to score; get back class probabilities or point predictions. ChakraTab fits in-context on your table at request time: no training pipeline, no hyperparameters to search, calibrated probabilities for classification and quantile-ready predictions for regression. It is listed on TabArena as **Chakra-Tab**.

- Website: https://yhatlabs.com
- Request API access: https://yhatlabs.com/models/access
- Time-series model: [yhatlabs/ChakraTS](https://huggingface.co/yhatlabs/ChakraTS)
- Leaderboards: [TabArena](https://huggingface.co/spaces/TabArena/leaderboard) (submission under review, [PR #643](https://github.com/autogluon/tabarena/pull/643))

## Getting started

Access is by API key, issued on request. One call fits on the training rows and predicts the test rows:

```bash
curl -X POST https://api.yhatlabs.com/v1/tabular/predict \
  -H "Authorization: Bearer $YHAT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "op": "fit_predict",
        "target": "churned",
        "preset": "medium",
        "train": [
          {"tenure": 12, "plan": "pro", "tickets": 3, "churned": 1},
          {"tenure": 48, "plan": "basic", "tickets": 0, "churned": 0}
        ],
        "test": [
          {"tenure": 6, "plan": "pro", "tickets": 5},
          {"tenure": 30, "plan": "basic", "tickets": 1}
        ]
      }'
```

Response (abridged):

```json
{
  "classes": ["0", "1"],
  "probabilities": [[0.21, 0.79], [0.88, 0.12]],
  "predictions": ["1", "0"],
  "fit": {"problem_type": "binary", "eval_metric": "log_loss", "n_rows": 2, "n_features": 3, "fit_s": 4.1}
}
```

Large tables travel as base64 parquet (`{"format": "parquet_b64", "bytes": "..."}`) in place of the row lists. A scikit-learn-style Python client (`fit` / `predict` / `predict_proba`) is in preparation.

### Input and output contract

| Field | Meaning |
|---|---|
| `train` | rows as objects (or a parquet payload); numeric, categorical and text columns; `null` marks a missing value |
| `target` | the column of `train` to predict |
| `test` | rows with the same columns minus the target |
| `problem_type` | optional: `binary`, `multiclass` or `regression`; inferred when omitted |
| `eval_metric` | optional: `log_loss`, `roc_auc` or `rmse`; sets what the internal validation optimises |
| `preset` | `fast` (single holdout, seconds), `medium` (8-fold bagging, default), `full` (8-fold bagging with fine-tuning, minutes to hours) |
| `time_limit` | optional fit budget in seconds |
| response | `probabilities` and `classes` for classification, `predictions` always; `fit` reports what was fitted |

## Capabilities

- **Any table**: binary, multiclass and regression; numeric, categorical, text and missing values handled without preprocessing.
- **Calibrated probabilities**: classification is scored and selected on log loss, so the probabilities are usable as they are.
- **Presets trade time for accuracy**: `fast` answers in seconds for interactive use; `medium` is the default for production; `full` is the benchmark configuration.
- **No training data retained**: the model is fitted in-context per request and discarded afterwards unless you ask to keep it.

## Evaluation

[TabArena](https://tabarena.ai) is the living benchmark for tabular machine learning (51 datasets, Elo against every entrant). The numbers below are TabArena-Lite (the first split of every dataset) with the official pipeline, scored by TabArena's own Elo machinery against the 98 leaderboard entries at submission time (October 2026); the maintainers' full run follows the review.

| Entry | Elo | Mean rank | Win rate |
|---|---|---|---|
| LimiX-2 (non-commercial) | 1919 | 7.0 | 0.94 |
| TabPFN-3.5 | 1858 | 9.0 | 0.92 |
| TabFM+ (non-commercial) | 1834 | 9.85 | 0.909 |
| **ChakraTab `full`** | **1834** | **9.86** | **0.909** |
| AutoGluon 1.6 (non-commercial, 4h) | 1806 | 11.0 | 0.90 |
| Mitra-v2 | 1755 | 13.1 | 0.88 |
| **ChakraTab `medium`** | **1745** | **13.6** | **0.87** |
| AutoGluon 1.6 (extreme, 4h) | 1746 | 13.5 | 0.87 |

Per-split results and the entry-point script are in the [submission](https://github.com/autogluon/tabarena/pull/643) and its [release of raw results](https://github.com/yhatlabs/tabarena/releases/tag/tabarena-lite-results-2026-10-05).

## Approach

ChakraTab combines several pretrained tabular foundation models. Each fits in-context on your table; their predictions are validated out-of-fold on your rows and combined with weights chosen on that validation, so the combination is specific to each table. Nothing is learned across customers or across requests.

## Intended use

- **Risk and churn:** default, fraud, churn and conversion scoring with calibrated probabilities.
- **Operations:** failure, delay and demand classification from sensor and ERP tables.
- **Pricing and valuation:** regression on property, vehicle, insurance and product tables.
- **Analytics:** a strong baseline for any new tabular problem before a bespoke model is justified, in one call.
- **Backtesting:** evaluate on your own held-out rows before adopting a model, using the same endpoint.

## Not intended use

- Tables with more than a few hundred thousand rows per fit; sample or aggregate first.
- Time-series forecasting; use [ChakraTS](https://huggingface.co/yhatlabs/ChakraTS).
- Text-heavy records where the text, not the table, carries the signal.

## Data handling

Tables sent to the API are used only to produce the response. They are not stored beyond the request, not used for training and not shared with third parties. Request metadata (timestamps, sizes, latency) is logged for operations.

## Terms

Use of the API is governed by the [YHat Labs API terms](https://yhatlabs.com). ChakraTab is available for commercial use under those terms; enterprise and on-premises arrangements on request.

## Citation

```bibtex
@misc{chakratab2026,
  title  = {ChakraTab: in-context tabular prediction as a service},
  author = {YHat Labs},
  year   = {2026},
  url    = {https://huggingface.co/yhatlabs/ChakraTab}
}
```

## Contact

founders@yhatlabs.com · https://yhatlabs.com