---
configs:
- config_name: fluxreason
  data_files:
  - split: train
    path: "fluxreason/train-*.parquet"
  default: true
- config_name: gptedit
  data_files:
  - split: train
    path: "gptedit/train-*.parquet"
- config_name: imagenet22k
  data_files:
  - split: train
    path: "imagenet22k/train-*.parquet"
- config_name: inaturalist
  data_files:
  - split: train
    path: "inaturalist/train-*.parquet"
- config_name: megalith10m
  data_files:
  - split: train
    path: "megalith10m/train-*.parquet"
- config_name: midjourneyv6
  data_files:
  - split: train
    path: "midjourneyv6/train-*.parquet"
- config_name: pexels
  data_files:
  - split: train
    path: "pexels/train-*.parquet"
- config_name: places365-challenge2016
  data_files:
  - split: train
    path: "places365-challenge2016/train-*.parquet"
- config_name: redcaps
  data_files:
  - split: train
    path: "redcaps/train-*.parquet"
- config_name: rendered_text
  data_files:
  - split: train
    path: "rendered_text/train-*.parquet"
- config_name: textatlas
  data_files:
  - split: train
    path: "textatlas/train-*.parquet"
- config_name: yfcc
  data_files:
  - split: train
    path: "yfcc/train-*.parquet"
task_categories:
- text-to-image
size_categories:
- 100M<n<1B
---

# i1-captions-selected: one complete caption per image

A derivative of [`zlab-princeton/i1-captions`](https://huggingface.co/datasets/zlab-princeton/i1-captions)
that (a) **repairs** the captions that the original pipeline cut off mid-sentence, and
(b) **collapses** the up-to-five alternative captions per image down to a single chosen one.

The original dataset stores `caption1`..`caption5` and the i1 recipe samples one at random per
training iteration. Here that choice is made once, up front, and the chosen caption is
guaranteed to be a complete sentence — so the data can be consumed directly as `(key, caption)`.

## 1. What changed vs. the original

**Repaired truncation.** 1,152,605 rows (0.69%) had *no* complete caption in any column — an
artifact of the original captioner running with a 256-token generation cap. Those images were
re-captioned from source with the i1 recipe: **`Qwen/Qwen3-VL-30B-A3B-Instruct-FP8`** in FP8,
the same prompts (including the TextAtlas ground-truth-annotation prompt and the RenderedText
OCR-line prompt), images RGB → shorter-edge-512 → center-crop, and a raised token budget
(640 → 1536, and 3072 for the OCR-transcription-heavy `textatlas`, where 640 was badly
insufficient: repaired captions there average 293 words vs ~160 under the old cap).

**One caption per row.** The `caption1`..`caption5` columns are replaced by a single `caption`.

## 2. Layout

Same structure and `key` values as the original: one directory per subset, `train-000NN.parquet`
shards of 1M rows.

```
<subset>/train-000NN.parquet     # columns: key (string), caption (string)
```

| subset | rows | vs. original |
| --- | ---: | --- |
| `yfcc` | 97,942,438 | −2,848 |
| `imagenet22k` | 13,673,542 | −2 |
| `rendered_text` | 11,977,780 | −36 |
| `megalith10m` | 9,393,285 | −686 |
| `places365-challenge2016` | 7,220,913 | −684 |
| `fluxreason` | 5,890,279 | = |
| `textatlas` | 5,396,767 | −123 |
| `redcaps` | 4,817,430 | −1 |
| `inaturalist` | 4,812,852 | −691 |
| `pexels` | 2,810,634 | = |
| `gptedit` | 1,553,575 | = |
| `midjourneyv6` | 1,240,185 | = |

**Total: 166,729,680 rows** — 5,071 fewer than the original 166,734,751 (0.003%).

Those 5,071 rows were **dropped**, not left broken. Their source images are dead links, so no
clean caption exists for them anywhere and repeated resampling could not produce one. Breakdown
of what was dropped: 4,117 captions under 25 words (on inspection these are not merely terse but
collapsed — broken grammar, language mixing, and outright non-answers like "I cannot describe an
image that was not provided"), 918 that could not be made to end on a sentence boundary, 21
repetition loops, and 15 assistant refusals.

```python
from datasets import load_dataset
ds = load_dataset("<your-namespace>/i1-captions-selected", "yfcc", split="train")
```

## 3. How the caption was chosen

1. **Repaired rows win** — used verbatim.
2. **Otherwise the best of `caption1`..`caption5`.** The five columns are statistically
   interchangeable (measured: near-identical completion rates and length distributions), so
   there is no "better column"; candidates are filtered on content. A candidate must end on
   sentence-final punctuation, contain no structural junk (` ``` `, `bbox_2d`, `"image_path"`,
   `<html>`, …), not open with a refusal, contain no repetition loop, and be ≥25 words. Among
   survivors the longest wins — the i1 paper finds longer caption sets train better.
3. **Salvage** — if nothing passes, every candidate is trimmed to its last complete sentence and
   the richest survivor is used (0.08% of rows).
4. **Drop** — if even that fails, the row is omitted (5,071 rows total, 0.003%).

A "repetition loop" means the captioner degenerated into repeating itself (e.g. one caption had
`ME` 211 times in 421 words). The detector is deliberately strict — a single token appearing
≥50 times *and* making up >25% of the caption, or a 5-gram repeated ≥20 times — because a looser
rule wrongly flags correct OCR transcription of images that genuinely contain long runs of a
character.

`imagenet22k`'s extra columns (`qwen2vl_2b`, `qwen2.5vl_3b`, `qwen3vl_2b`, `qwen3vl_4b`,
`short`, `no_center_crop`) are **excluded** — they are ablation artifacts from weaker captioners
and different preprocessing/prompt settings, not the captions used for i1's final training.

## 4. Quality

Audited over all 166.7M rows, every subset: **0 incomplete, 0 repetition loops, 0 refusals,
0 empty, 0 under 25 words.** Mean caption length 128–174 words depending on subset.

## 5. One thing to know before training

i1 right-truncates text-encoder input to **256 tokens**. The repaired captions are longer than
the original corpus, which was implicitly bounded by the old 256-token generation cap: measured
with the `google/t5gemma-2b-2b-ul2` tokenizer, the original `yfcc` `caption1` is median 172
tokens with 0.1% over 256, whereas the repaired ones are median 246 with ~41% over 256. Those
rows will be cut at the encoder. Trimming to the last complete sentence within 256 tokens is a
lossless offline fix if that matters for your run; it is deliberately **not** applied here.

## 6. Citation

The underlying dataset and recipe are from:

```bibtex
@article{zeng2026i1,
  title={i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models},
  author={Zeng, Boya and Luo, Tianze and Pu, Shu and Shen, Jucheng and Lu, Taiming and Sarch, Gabriel and Liu, Zhuang},
  journal={arXiv preprint arXiv:2606.11289},
  year={2026}
}
```