vanloc1808/pico-banana-smolvlm-format-with-rejected-answer

pico-banana-smolvlm-format-with-rejected-answer

Balanced image-level tampering detection dataset in SmolVLM-style format with chosen/rejected answer pairs, derived from the pico-banana MCQ pipeline. Suitable for preference learning (e.g. DPO) and RLHF-style training.

Dataset overview

Same as vanloc1808/pico-banana-smolvlm-format, but each example includes a rejected_answer field: the answer from the counterpart sample (same edited/original image pair, opposite image_used_as_main). This allows training models to prefer the correct (chosen) explanation over the incorrect (rejected) one.

• —Each example: one image, one question, chosen answer (answer), rejected answer (rejected_answer), and binary label (0 = original, 1 = edited).
• —Images stored as datasets.Image; downloaded/decoded automatically.
• —Produced by add_rejected_answer.py from the balanced JSONL, then pushed to the Hub.

Data fields

• —image (datasets.Image): The image (original or edited, per label).
• —question (string): Fixed prompt (same as base dataset).
• —answer (string): Chosen explanation (correct for this image).
• —rejected_answer (string): Rejected explanation (from counterpart sample with opposite label).
• —label (int64): 0 = original, 1 = edited/tampered.

Splits

train and validation (default 90/10 split, seed 42), same as base dataset.

Usage

from datasets import load_dataset

repo_id = "vanloc1808/pico-banana-smolvlm-format-with-rejected-answer"
ds = load_dataset(repo_id)
ex = ds["train"][0]
image = ex["image"]
question = ex["question"]
chosen = ex["answer"]
rejected = ex["rejected_answer"]
label = ex["label"]

License and attribution

Same as the base pico-banana dataset. Refer to upstream sources for licensing.