EditPPT/editppt-vision-validator-lora

EditPPT Vision Validator (LoRA on Qwen3.5-9B)

A LoRA adapter that checks whether an edit to a PowerPoint slide newly introduced a severe visual defect, and if so, returns a repair as tool calls — in one generation. It serves as the vision validator of EditPPT.

Given the slide rendered before and after an edit, plus the edit instruction, the tool calls that made the edit and the slide's shape JSON, the model returns:

{"HasCriticalIssues": "Yes",
 "Issues": [{"IssueType": "ELEMENT_OVERLAP", "AffectedShapeIDs": [2],
             "TechnicalDiagnosis": "textbox 2 overlaps the caption below it by about 20 pt"}],
 "ToolCalls": [{"name": "adjust_layout", "arguments": {"slide_number": 3, "shape_id": 2, "top": 470}}]}

• Defect types: CONTENT_OFF_SLIDE, ELEMENT_OVERLAP, LOW_CONTRAST.
• Repair tools: set_shape_font_size, adjust_layout, set_text_style. Values are absolute targets in slide points.
• Defects already present in the BEFORE image are treated as the deck's own design and are not reported.

Usage

See inference_example.py and prompt_template.txt.

• Message order: [image BEFORE, image AFTER, prompt text].
• Images: 960x540 (min_pixels = max_pixels = 960*540). One slide point = one pixel for a 16:9 deck.
• Greedy decoding, default chat template.

Training

Base	Qwen3.5-9B, bf16
Adapter	LoRA, r 16, alpha 32, on the language model and the vision encoder
Data	7,260 training samples (plus 374 for validation) with human-confirmed labels, from publicly licensed decks (CC0 / CC-BY)

Evaluation

Held-out set of 750 samples from decks not seen in training:

Metric	Value
Verdict accuracy, edits without a repair (n=497)	0.932 (base model: 0.60)
Verdict accuracy, all 750 samples	0.955
Defect precision / recall / F1, all 750 samples	0.963 / 0.954 / 0.959
Repair targets the right shapes with a permitted tool, all 750 samples	0.948

License

Apache-2.0, following the base model.