README.md

---
license: cc-by-4.0
pretty_name: Auto-Fill Benchmark
language:
  - en
task_categories:
  - table-question-answering
  - text-generation
tags:
  - tabular
  - spreadsheets
  - data-cleaning
  - missing-value-imputation
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: test
        path: index.jsonl
---

# Auto-Fill Benchmark

Benchmark for **predicting missing cell values in real-world tables**, introduced in
*Auto-Fill: Learning to Predict Missing Values Accurately with Specialist Language Models*
(PVLDB 19(11), 2026 — [arXiv:2607.19847](https://arxiv.org/abs/2607.19847)).
Each case is a real table in which exactly one cell is replaced by `[MISSING]`, together with the ground-truth value.

- **Code:** https://github.com/lyrain2001/auto-fill
- **Models:** [Auto-Fill-Qwen3-8B-Knowledge](https://huggingface.co/lyrain2001/Auto-Fill-Qwen3-8B-Knowledge) ·
  [Auto-Fill-Qwen3-8B-Reasoning](https://huggingface.co/lyrain2001/Auto-Fill-Qwen3-8B-Reasoning) ·
  [Auto-Fill-Qwen3-8B-Coding](https://huggingface.co/lyrain2001/Auto-Fill-Qwen3-8B-Coding)

## What is included

9 of the paper's 11 benchmark datasets, 200 cases each (1,800 cases in total):

| Dataset | Split | Source |
|---|---|---|
| `Pub-XLS` | ID | Relational tables parsed from `.xlsx` spreadsheets crawled from a search-engine index |
| `Pub-BI` | ID | Relational tables extracted from public business-intelligence (BI) models |
| `Pub-Wiki` | ID | Tables from a recent Wikipedia snapshot |
| `Gov-CSV` | ID | CSV files crawled from nationalarchives.gov.uk |
| `Git-Parquet` | ID | Parquet files crawled from GitHub |
| `Pub-Web` | OOD | General web tables extracted from public HTML pages |
| `Rel-AR` | OOD | Tables with column-level arithmetic relationships (e.g. `TOTAL = PRICE * QUANTITY`) |
| `Rel-FD` | OOD | Tables with real functional dependencies (e.g. `ProductKey → ProductName`) |
| `Rel-ST` | OOD | Tables with string-based relationships (e.g. full name = first name + last name) |

ID / OOD = in-distribution / out-of-distribution with respect to the table sources used to train the Auto-Fill specialists (test tables are disjoint from training tables). For the `Rel-*` datasets the relationship is **not** given to the model; it has to be inferred from the table.

### Not released: `Ent-CSV` and `Ent-XLS`

The paper's two remaining benchmarks — **Ent-CSV** (200 CSV files from a large enterprise's data lake) and
**Ent-XLS** (200 enterprise `.xlsx` spreadsheets) — are proprietary enterprise data behind a corporate firewall
and cannot be released. Results on these two datasets therefore cannot be reproduced from this release; all
other numbers in the paper can.

## Layout

```
sample200/
  <Dataset>/
    case_<i>/
      data.csv              # the table; exactly one cell is "[MISSING]"
      info.json             # position of the masked cell + ground truth
      ground_truth.csv      # Rel-AR, Rel-FD only: annotated column relationships (see below)
      ground_truth_sem.csv  # Rel-AR only
index.jsonl                 # one row per case (drives the dataset viewer)
```

`info.json`, e.g.
```json
{"row_idx": 5, "col_idx": 0, "col_name": "FID", "label": "Parks.fid--c59f932_14cc29bd18f_-dfc", "output": "{\"value\": \"Parks.fid--c59f932_14cc29bd18f_-dfc\"}"}
```
`row_idx` / `col_idx` are 0-based (the row index does not count the header); `label` is the ground-truth value as a string; `output` (where present) is the same value wrapped as JSON.

Supplementary relationship annotations (not used by the evaluation code):
- `Rel-AR/ground_truth_sem.csv` — `formula` in column-header names (e.g. `TOTAL=PRICE*QUANTITY`); `ground_truth.csv` is the same with Excel column letters (`A` = first column). `sample_type` P/N = relationship holds / has violations.
- `Rel-FD/ground_truth.csv` — `left_col,right_col,sample_type`: candidate functional dependencies `left_col → right_col` (P = holds, N = violated).
- `Rel-ST` — the relationship annotations refer to columns of an auxiliary joined table that is not part of `data.csv`, so they are not included.

`index.jsonl` fields: `dataset, case, distribution, row_idx, col_idx, col_name, label, n_rows, n_cols, table_csv`.

## Usage

Download with the layout the code expects:
```bash
hf download lyrain2001/Auto-Fill-Benchmark --repo-type dataset --local-dir Auto-Fill-Benchmark
```

Run one specialist on one dataset and evaluate (from the [code repository](https://github.com/lyrain2001/auto-fill)):
```bash
python inference/run_benchmark.py --mode knowledge \
    --model_path lyrain2001/Auto-Fill-Qwen3-8B-Knowledge \
    --dataset Gov-CSV --benchmark Auto-Fill-Benchmark/sample200 --gpu_ids 0
python evaluation/evaluate_single_specialist.py \
    --results_dir results/benchmark/benchmark_knowledge --benchmark Auto-Fill-Benchmark/sample200
```

Or load the index with 🤗 Datasets:
```python
from datasets import load_dataset
ds = load_dataset("lyrain2001/Auto-Fill-Benchmark", split="test")
```

## Evaluation

A prediction counts as correct if it matches `label` under `autofill.utils.is_correct` (normalized string comparison
with numeric tolerance and date/time parsing). Because models may abstain, the primary metric is
**Recall@Precision=0.9 (R@P90)** — the fraction of cells filled correctly when a method only answers where its
confidence is high enough to keep precision ≥ 0.9 — together with pAUPRC.

Auto-Fill with the three Qwen3-8B specialists and the calibrated ensemble (R@P90 / pAUPRC, from the paper):

| Pub-XLS | Pub-BI | Pub-Wiki | Gov-CSV | Git-Parquet | Ent-CSV* | Ent-XLS* | Pub-Web | Rel-AR | Rel-FD | Rel-ST | Mean |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.53 / 0.48 | 0.62 / 0.61 | 0.29 / 0.28 | 0.50 / 0.49 | 0.59 / 0.58 | 0.59 / 0.56 | 0.66 / 0.64 | 0.28 / 0.23 | 0.99 / 0.99 | 0.89 / 0.89 | 1.00 / 0.99 | 0.63 / 0.61 |

\* not released (see above).

## License

The annotations in this release (masked cells, labels, relationship files, `index.jsonl`) are provided under
CC BY 4.0. The tables themselves come from public sources (Wikipedia, nationalarchives.gov.uk, public GitHub
repositories, public web pages, public BI models and spreadsheets) and remain subject to the terms of their
original sources. Intended for research use.

## Citation

```bibtex
@article{liu2026autofill,
  title={Auto-Fill: Learning to Predict Missing Values Accurately with Specialist Language Models},
  author={Liu, Yurong and He, Yeye and Dong, Haoyu and Xing, Junjie and Han, Shi and Zhang, Dongmei and Chaudhuri, Surajit},
  journal={Proceedings of the VLDB Endowment},
  volume={19},
  number={11},
  pages={3160--3173},
  year={2026}
}
```