experiment/knowledge_editing/README.md

Knowledge Editing Baselines

Compare our OVERTONE LoRA method against standard knowledge model editing (KME) methods from EasyEdit on the same hallucination suppression task.

Problem Formulation

The Edit (Captioning)

Each edit instance is a concrete captioning correction:

Image_42 + "Describe this image."
  old:  "A bathroom with a toilet, sink, and mirror"    ← hallucinated
  new:  "A bathroom with a sink and mirror"              ← toilet removed

The target is the model's own caption with toilet mentions surgically removed. This is a minimal edit — the model's style, vocabulary, and all correct content are preserved. Only the hallucinated part changes.

Why Captioning (not VQA)

• Captioning is where the hallucination manifests in practice
• Each edit is a concrete (input, old_output → new_output) triple — exactly what KME methods are designed for
• The target preserves the model's own distribution (minimal intervention)
• All methods are evaluated on the same metric: toilet mention rate in captions

Mapping to Standard KME

KME Concept	Standard (text-only)	Our Setting
Subject s	"Eiffel Tower"	bathroom-without-toilet image
Relation r	"located in"	"Describe this image."
Old object o	"Paris"	"...a toilet, sink, and mirror"
New object o*	"London"	"...a sink and mirror"
Rephrase (text)	"Where is the Eiffel Tower?"	"What objects are in this picture?"
Rephrase (visual)	—	Different bathroom-no-toilet image
Locality (text)	Unrelated facts	"What is the capital of France?"
Locality (vision)	—	Bathroom WITH toilet (should still mention toilet)

Key Differences from Standard KME

1. Distributional subject — not one entity but a distribution of images
2. Suppression — target is "same minus toilet", not a fixed replacement
3. Visual conditioning — same token must be suppressed or preserved depending on what the image shows
4. Per-instance targets — each image has a different correct caption

Pipeline

Step 1: Build edit set (no GPU)    →  edit_set.json (structure, no targets)
Step 2: Generate targets (GPU)     →  edit_set.json (filled with captions)
Step 3: Run EasyEdit methods       →  edited models
Step 4: Evaluate all               →  comparison table

Step 1+2: Build Edit Set

# Structure only (fill targets on server later)
python -m experiment.knowledge_editing.build_edit_set \
    --csv CC3M-Dataset/bathroom_filter/bathroom_toilet_labels.csv \
    --image_dir CC3M-Dataset/cc3m_images/train \
    --output experiment/knowledge_editing/edit_set.json \
    --max_locality_per_category 50

# Generate targets on server (needs GPU)
python -m experiment.knowledge_editing.build_edit_set \
    --fill_targets experiment/knowledge_editing/edit_set.json \
    --model llava-hf/llava-1.5-7b-hf

# Or do both in one go
python -m experiment.knowledge_editing.build_edit_set \
    --csv ... --image_dir ... --output ... \
    --generate_targets --model llava-hf/llava-1.5-7b-hf

The target generation step:

1. Runs original LLaVA on each bathroom-no-toilet image
2. Gets the hallucinated caption ("A bathroom with a toilet and sink")
3. Removes toilet mentions ("A bathroom with a sink")
4. Checks if the result is still a valid caption (min 4 words)

Step 3: Run Baselines

PYTHONPATH="${PYTHONPATH}:./EasyEdit" \
python -m experiment.knowledge_editing.run_baselines \
    --edit_set experiment/knowledge_editing/edit_set.json \
    --methods wise grace lora \
    --n_edits 20 \
    --output_dir step4_ke_outputs

Step 4: Compare with Our Method

# Run OVERTONE LoRA
python -m experiment.training.finetune_lora_v3 --config experiment/lora_v3_config.json

# Evaluate (same pipeline for all methods)
python -m experiment.evaluation.validate \
    --model_type lora \
    --model_dir <path> \
    --inference_backend transformers \
    --mention_method keyword

File Reference

File	Purpose
build_edit_set.py	CSV → edit_set.json + target generation
llava15_compat.py	LLaVA-1.5 processor wrapper for EasyEdit
run_baselines.py	Run EasyEdit methods + save + evaluate
hparams/*.yaml	EasyEdit configs per method

Edit Set Format

{
    "edit_instances": {
        "train": [
            {
                "image_id": "12345",
                "image_path": "...",
                "original_caption": "A bathroom with a toilet, sink, and mirror",
                "target": "A bathroom with a sink and mirror",
                "had_toilet": true,
                "is_usable": true
            }
        ]
    },
    "locality_instances": {
        "bathroom_with_toilet": [
            {
                "image_id": "67890",
                "original_caption": "A bathroom with a white toilet and bathtub"
            }
        ]
    }
}