Overdog/MFRL_Data

MFRL Data

Instruction-based image edits that went wrong, each paired with what should have happened and how to fix it, plus the MFRL-Bench test set.

• default / train: 50,642 records. Each has a source image, an edit instruction, a ground-truth result, a failed result with an obvious error, and feedback that names the problem and then gives a corrective instruction.
• default / test: 3,000 records, the edit tasks of MFRL-Bench. Each has the source, the instruction and the benchmark's reference result. 2,738 of them also carry a failed image and feedback built the same way as the training data. The others have none (failed_image and feedback are null), because no obvious failure was found.
• judge / test: 2,000 images, the judge track of MFRL-Bench: edited images for the test tasks, half labeled success and half fail, each with a written reason.

No overlap between train and test. The 2,738 test tasks that also had a training record were removed from train, and so were 49 more records that show the same photo as a test task (same Open Images ID, same source file, same OmniEdit scene, or a 256-bit image-hash distance of 25 or less).

All images are 1024×768 PNG and are embedded in the Parquet files.

Loading

from datasets import load_dataset

train = load_dataset("Overdog/MFRL_Data", split="train")
test = load_dataset("Overdog/MFRL_Data", split="test")
judge = load_dataset("Overdog/MFRL_Data", "judge", split="test")

r = train[0]
r["source_image"], r["target_image"], r["failed_image"]   # PIL images
r["instruction"], r["feedback"]

Fields (default config)

Field	Type	Meaning
id	string	Record ID; the upstream sample it was built from
edit_type	string	One of the 9 content-edit classes
instruction	string	Edit instruction given to the editor
source_image	image (PNG)	Image to edit
target_image	image (PNG)	Ground-truth edit result
failed_image	image (PNG)	Edit result with an obvious error
feedback	string	Problem in the failed image, then the corrective instruction (at most 150 words)
feedback_problem	string	First part of feedback: what is wrong and where
feedback_correction	string	Second part of feedback: what the editor should do instead
failed_image_origin	string	original_candidate (kept from the candidate pool that produced target_image) or regenerated (one of three new samples; the model is in failed_image_model)
failed_image_model	string	Model that generated failed_image
failed_image_seed	int64	Seed of failed_image
target_image_model	string	Model that generated target_image
target_human_confirmed	bool	Whether a person confirmed target_image (otherwise chosen by an automatic judge)
feedback_model	string	Model that judged failed_image and wrote feedback
source_dataset	string	Dataset the source image and instruction come from
source_subset	string	Subset within source_dataset
image_key	string	Source image ID; split train/test by this column
split_origin	string	ours (train, or a test task with our quad) or mfrl_bench (test task without a quad)
mfrl_id	string	MFRL-Bench task id (test split only)
mfrl_category	string	MFRL-Bench category (test split only; the 4 text classes are one category, Text)

Fields (judge config)

Field	Type	Meaning
id	string	MFRL-Bench judge item id
edit_id	string	MFRL-Bench edit task the image was made for
category	string	MFRL-Bench category
subcategory	string	Text sub-type, else null
instruction	string	Edit instruction
source_image	image (PNG)	Image to edit
candidate_image	image (PNG)	Edited image to judge
label	string	success or fail
reason	string	Labeling judge reason (Claude Haiku 5.5)
confidence	string	Labeling judge confidence
generator	string	Model that produced the candidate
label_source	string	How the label was obtained

Counts

Train by edit type:

Edit type	Records
Change Object Appearance	21,753
Replace Object	8,646
Add Object	7,648
Remove Object	4,919
Complex Edit	2,595
Add Text	1,925
Move Object	1,405
Replace Text	942
Change Text Style	465
Remove Text	344

Test by MFRL-Bench category:

Category	Tasks
Add Object	500
Change Object Appearance	500
Complex Edit	500
Remove Object	500
Replace Object	500
Text	400
Move Object	100

Train by model of the failed image:

Model	Records
FLUX.2-klein-9B	30,923
FLUX.2-klein-4B	9,393
Qwen-Image-Edit-2511-Lightning	5,088
FLUX.3	5,059
Qwen-Image-2.1	179

Train by source dataset:

Source	Records
GPT-Image-Edit-1.5M	37,602
Pico-Banana-400K	8,753
Open Images (TextOCR)	3,244
Open Images	1,043

How it was built

1. Sources and ground truth. Source images and instructions come from GPT-Image-Edit-1.5M (OmniEdit subset), Pico-Banana-400K and Open Images V6 (with TextOCR for text edits). Several candidate edits were generated per source with FLUX.3 and Qwen-Image-2.1, and an automatic judge (Claude Haiku 5.5) chose a successful one as the target.
2. Failed image. If a failed candidate from that pool had an obvious error, meaning one a person notices within seconds at normal size, it was kept. Otherwise FLUX.2-klein-9B (4 steps) generated three new results and the judge picked the one with the most obvious error. If none had one, FLUX.2-klein-4B (4 steps) tried three more, and then Qwen-Image-Edit-2511 with the Lightning 4-step LoRA. If all nine had no obvious error, the record was dropped. Small shifts in color, tone or sharpness, and differences of taste on style edits, do not count as errors.
3. Feedback. The same judge call wrote the feedback: the problem first, then an imperative correction, in words rather than coordinates, at most 150 words. The judge was Claude Haiku 5.5 for most records and Claude Opus 5.5 for part of the text, move and complex-edit records; feedback_model names it per record.
4. Test set. MFRL-Bench tasks were selected where two independent Claude Haiku 5.5 judgments agreed on every candidate and on the best one, which is the reference. No Pico-Banana-400K image is in the test set.

Limitations

• Labels are automatic. No person reviewed the failed images, the feedback, the references or the judge labels.
• Generators differ. In train, most failed images come from FLUX.2-klein, while targets come from FLUX.3. A model trained to tell them apart can learn the generator instead (failed_image_model, target_image_model).
• Selection bias. Records where every tried model got the edit right are missing, so easy edits are underrepresented.
• The reference is one valid result, not the only one.

License

The data combines material under different terms; see LICENSE.md. In particular, records with source_dataset = Pico-Banana-400K are under CC BY-NC-ND 4.0 (non-commercial, no derivatives). All target_image, failed_image and candidate_image files are AI-generated.