sref/README.md

FreeStyle Dataset — sref split

This subfolder is the cref+sref+target triplet export of the cref_sref_oneig_filter_part1..part5 vault sources used by the VGO project. The layout is intentionally simple so that downstream readers can use the dataset with nothing more than Python's standard library + Pillow + pandas — DuckDB and the internal vault package are not required.

Contents

sref/
├── README.md              # this file
├── pairs.csv              # one row per (sequence_id, choice_id) — fully self-contained
├── metadata.jsonl         # one row per sequence_id, preserves the original vault structure
├── export_summary.json    # counts and export-time metadata
└── images/<uu>/<image_uuid>.<ext>
                           # deduplicated image pool. <uu> is the first two hex chars of the uuid.

• Exported sources: 5
• cref_sref_oneig_filter_part1
• cref_sref_oneig_filter_part2
• cref_sref_oneig_filter_part3
• cref_sref_oneig_filter_part4
• cref_sref_oneig_filter_part5
• Unique vault sequences exported: 417,761
• Rows in pairs.csv (= sequences × choices): 3,342,076
• Unique images in images/: 1,238,604
• Export wall-clock: started 2026-06-11T08:11:57.855884Z, finished 2026-06-11T08:33:59.192685Z

Schema — pairs.csv

Column	Description
sequence_id	Vault sequence UUID. Eight rows share the same id, one per choice.
choice_id	0..7. Index of the instruction × caption × language combination.
source	One of cref_sref_oneig_filter_part{1..5}.
task_type	Always edit for this export.
cref_image / sref_image / target_image	Relative path inside this folder, e.g. images/ab/ab12….webp. Open with PIL.
cref_width / cref_height / sref_* / target_*	Pixel dimensions reported by the vault.
instruction / instruction_index	Instruction text + its vault-internal key (sample_instruction_en_123 etc).
target_caption / target_caption_index	Target image caption + its vault key (captions/scene_3 etc).
content_trigger_words / style_trigger_words	Trigger phrases attached to the content/style references.
message_formatter	The original modality template, always <image_0><image_1><text_0><text_1><image_2>.
require_loss_json	JSON list aligned with the formatter modalities — [false, false, false, false, true] for edit.
uri	Vault-side original uri string.
create_time	ISO-8601 timestamp recorded in the vault.

Loading with pandas (no DuckDB needed)

import pandas as pd
from PIL import Image
from pathlib import Path

root = Path("/path/to/sref")          # local clone of this folder
df = pd.read_csv(root / "pairs.csv")

row = df.iloc[0]
cref = Image.open(root / row["cref_image"])
sref = Image.open(root / row["sref_image"])
target = Image.open(root / row["target_image"])
prompt = row["instruction"]
caption = row["target_caption"]

To keep only English instructions:

en_df = df[df["instruction_index"].str.endswith("_en_123")]

To deduplicate to one row per sequence (e.g. for visualizing pairs only):

unique = df.drop_duplicates(subset="sequence_id")

Loading with datasets

from datasets import load_dataset

ds = load_dataset(
    "Blue2Giant/FreeStyle_Dataset",
    data_files="sref/pairs.csv",
    split="train",
)

Image columns are stored as relative paths inside this folder. After cloning the repo (or using huggingface_hub.snapshot_download), resolve them against the local snapshot root.

Schema — metadata.jsonl (optional, for power users)

Each line is a JSON object preserving the original vault structure for a single sequence_id. Useful if you need to:

• Reconstruct the eight sequence_choices exactly as the trainer saw them.
• Access the auxiliary texts that are not surfaced in pairs.csv, such as captions/scene_1, captions/scene_2, target_caption, and the per-language variants.

{
  "sequence_id": "...",
  "source": "cref_sref_oneig_filter_part1",
  "task_type": "edit",
  "uri": "...",
  "create_time": "2026-04-26T...",
  "images": [{"id": "...", "index": "cref_0", "width": 1024, "height": 1024, "file_name": "images/.../....webp"}, ...],
  "texts": [{"id": "...", "index": "...", "language": null, "content": "..."}, ...],
  "choices": [{"choice_id": 0, "message_formatter": "<image_0><image_1><text_0><text_1><image_2>", "image_indices": ["cref_0", "sref_0", "target"], "text_indices": ["sample_instruction_cn_123", "captions/scene_3_en"], "require_loss": [false, false, false, false, true]}, ...],
  "choices_weights": []
}

Image dedup

Each image is stored once under its vault UUID, regardless of how many sequences reference it. pairs.csv rows for different sequences therefore reuse the same cref_image / sref_image paths whenever a cref/sref appears in multiple pairs — this is the intended dedup behaviour and not a bug.

License & provenance

These triplets were assembled by sampling cref/sref references from internal collections; please refer to the parent repository's top-level dataset card for licensing.