Pico Banana Smolvlm Format With Rejected Answer

A balanced image-level tampering detection dataset in SmolVLM-style format. It includes chosen and rejected answer pairs derived from the pico-banana MCQ pipeline, suitable for preference learning and RLHF-style training. The dataset was created by author vanloc1808 and was last updated on February 15, 2026.

Use Cases

• Training Direct Preference Optimization (DPO) models based on the chosen/rejected answer pairs.
• Conducting RLHF-style training for multimodal models based on the structured preference data.
• Fine-tuning visual language models for image tampering detection based on the dataset's format.

Strengths

• Dataset is structured for preference learning with explicit chosen and rejected answer pairs.
• Derived from a described MCQ pipeline, suggesting a systematic creation process.
• Last updated on February 15, 2026, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

vanloc1808 on Hugging Face, derived from the pico-banana MCQ pipeline.

Collection Method

Derived from the pico-banana MCQ pipeline; specific method not detailed.

Freshness

Last updated 2026-02-15 15:58:32.