---
dataset_info:
features:
- name: image
dtype: image
- name: Text
dtype: string
- name: Label
dtype: string
- name: Paradigm
dtype: string
- name: Incident_based
dtype: string
- name: Humor
dtype: string
- name: Metaphor
dtype: string
- name: Metaphor Object
dtype: string
- name: Metaphor Explanation
dtype: string
- name: Misogynistic intensity
dtype: string
splits:
- name: train
num_bytes: 179710931.0
num_examples: 1500
- name: test
num_bytes: 60020368.0
num_examples: 500
download_size: 242960052
dataset_size: 239731299.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---
# BANMIME:Misogyny Detection with Metaphor Explanation on Bangla Memes
[](https://aclanthology.org/)
[](https://github.com/Ayon128/BANMIME)
[](https://www.kaggle.com/datasets/mdayonmia1804128/misogyny-meme-dataset)
[Md. Ayon Mia*](https://github.com/Ayon128),
[AKM Moshiur Rahman*](https://huggingface.co/pltops),
[Khadiza Sultana Sayma*](https://github.com/Khadiza13),
[Md Fahim*](https://github.com/md-fahim/),
[Md Tahmid Hasan Fuad](),
[Muhammad Ibrahim Khan](), and
[AKM Mahbubur Rahman]().
## Dataset Overview
- Misogyny detection (binary) and category identification (multi-class) dataset on 2,000 Bangla meme samples.
- Four misogyny categories: Stereotype (30.6%), Objectification (25.9%), Shaming (38.0%), and Violence (5.4%).
- Sourced from 3 social media platforms: Facebook (1,300), Instagram (450), and Reddit (250).
- Each sample includes misogyny labels, humor types, metaphor localization, and detailed human-written explanations.
- High inter-annotator agreement with Cohen's Îș = 0.74.
## Data Format
| **Column Title** | **Description** |
|-------------------------|---------------------------------------------------------------------------------|
| `meme_id` | Unique identifier for each meme entry. |
| `image_path` | Path to the meme image file. |
| `text_content` | Extracted text content from the meme. |
| `misogyny_label` | Primary misogyny category (Stereotype / Objectification / Shaming / Violence). |
| `humor_type` | Type of humor used (Mockery / Sarcastic / Ironic / Satirical / Other). |
| `metaphor_localization`| Location of metaphor (Text / Image / Both). |
| `metaphor_object` | Object used in metaphorical expression. |
| `misogyny_intensity` | Intensity level (High / Moderate / Low). |
| `meme_template` | Template category (Troll Face / Wojak / Doge / etc). |
| `explanation` | Detailed human-written explanation of misogynistic content. |
## đ Getting Started
To run the code, please check the scripts in the `Misogyny_Code/Evaluate` and `Misogyny_Code/LLaMA-Factory` folders.
---
## đ Vision-Language Model Performance
We evaluate baselines using both open-source and closed-source vision-language models across different prompting strategies and fine-tuning approaches.
### Zero-Shot Prompt
| Models | Sham | Stereo | Obj | Vio | Avg | BScore | LAVE | Avg Expl |
|------------------------|-------|--------|-------|-------|-------|--------|-------|----------|
| **Closed Source VLMs** | | | | | | | | |
| Gemini2.0 Flash | 34.97 | 52.55 | 65.77 | 72.41 | 56.43 | 86.70 | 35.00 | 60.85 |
| GPT-4o-mini | 26.26 | 61.33 | 60.33 | 77.42 | 56.34 | 87.26 | 27.20 | 57.23 |
| **Open Source VLMs** | | | | | | | | |
| Llama-3.2V 11B | 12.31 | 18.12 | 42.50 | 16.13 | 22.27 | 82.82 | 0.35 | 41.59 |
| Gemma-3-12B | 52.02 | 22.67 | 48.76 | 67.74 | 47.80 | 83.54 | 8.00 | 45.77 |
| Qwen2.5-VL 7B | 22.73 | 41.33 | 57.02 | 35.48 | 39.14 | 78.60 | 0.60 | 39.60 |
| Phi-3.5 | 23.23 | 67.33 | 25.62 | 6.45 | 30.66 | 82.30 | 0.23 | 41.27 |
| Phi-4 | 49.49 | 28.86 | 20.83 | 16.13 | 28.83 | 81.63 | 0.20 | 40.92 |
| LLaVA-1.5 7B | 16.92 | 39.60 | 27.50 | 19.35 | 25.84 | 82.40 | 0.22 | 41.31 |
| LLaVA-NeXT 7B | 29.59 | 25.68 | 16.53 | 16.67 | 22.12 | 79.25 | 0.45 | 39.85 |
### CoT Prompt
| Models | Sham | Stereo | Obj | Vio | Avg | BScore | LAVE | Avg Expl |
|------------------------|-------|--------|-------|-------|-------|--------|-------|----------|
| **Closed Source VLMs** | | | | | | | | |
| Gemini2.0 Flash | 48.11 | 64.23 | 47.75 | 55.17 | 53.82 | 86.96 | 31.60 | 59.28 |
| GPT-4o-mini | 22.54 | 80.69 | 44.55 | 33.33 | 45.28 | 87.20 | 22.94 | 55.07 |
| **Open Source VLMs** | | | | | | | | |
| Llama-3.2V 11B | 11.68 | 35.33 | 16.67 | 3.33 | 16.75 | 83.53 | 1.73 | 42.63 |
| Gemma-3-12B | 75.25 | 11.33 | 38.02 | 41.94 | 41.64 | 84.49 | 7.36 | 45.93 |
| Qwen2.5-VL 7B | 24.24 | 61.74 | 28.33 | 16.13 | 32.61 | 84.56 | 5.13 | 44.85 |
| Phi-4 | 10.71 | 59.46 | 14.17 | 2.45 | 22.09 | 84.41 | 1.29 | 42.85 |
| LLaVA-1.5 7B | 7.61 | 57.72 | 10.08 | 6.45 | 20.47 | 84.34 | 0.27 | 42.31 |
| LLaVA-NeXT 7B | 8.67 | 43.62 | 11.57 | 6.45 | 17.58 | 84.23 | 2.16 | 43.20 |
### LoRA (CoT) Fine-Tuning
| Models | Sham | Stereo | Obj | Vio | Avg | BScore | LAVE | Avg Expl |
|----------------|-------|--------|-------|-------|-------|--------|-------|----------|
| Llama-3.2V 11B | 32.83 | 65.33 | 52.89 | 48.39 | 49.86 | 86.01 | 8.00 | 47.01 |
| Gemma-3-12B | 16.16 | 76.67 | 49.59 | 64.52 | 51.74 | 86.16 | 13.80 | 49.98 |
| Qwen2.5-VL 7B | 44.95 | 59.33 | 48.76 | 51.61 | 51.16 | 85.36 | 5.00 | 45.18 |
| LLaVA-1.5 7B | 24.62 | 48.33 | 52.98 | 54.26 | 45.05 | 85.19 | 4.80 | 44.20 |
| Paligemma-2-10B| 8.08 | 61.33 | 42.98 | 61.29 | 43.42 | 81.80 | 2.00 | 41.90 |
---
## đ§ Fine-Tuning Strategies
We propose three distinct fine-tuning approaches:
- **Standard LoRA Fine-Tuning**: Basic fine-tuning with LoRA adaptation
- **Augmented Direct Prompt Fine-Tuning**: Enhanced training with data augmentation
- **Chain-of-Thought Fine-Tuning**: Structured reasoning supervision for improved metaphor understanding
### Comparative Performance
| Models | Sham | Stereo | Obj | Vio | Avg |
|---------------------|-------|--------|-------|-------|-------|
| **LLaMa-3.2V 11B** | | | | | |
| LoRAaug | 0.5 | 99.3 | 0.0 | 3.2 | 25.9 |
| LoRAstd | 11.6 | 61.3 | 49.6 | 61.3 | 46.0 |
| LoRACoT | 32.8 | 65.3 | 52.9 | 48.4 | 49.9 |
| **Gemma-3 12B** | | | | | |
| LoRAaug | 8.1 | 98.0 | 7.4 | 9.7 | 30.8 |
| LoRAstd | 10.1 | 70.7 | 49.6 | 48.4 | 44.7 |
| LoRACoT | 16.2 | 76.7 | 49.6 | 64.5 | 51.7 |
| **Qwen2.5-VL 7B** | | | | | |
| LoRAaug | 49.0 | 58.7 | 5.8 | 22.6 | 34.0 |
| LoRAstd | 47.0 | 59.0 | 27.3 | 37.1 | 42.6 |
| LoRACoT | 45.0 | 59.3 | 48.8 | 51.6 | 51.2 |
| **LLaVa-1.5 7B** | | | | | |
| LoRAaug | 13.1 | 84.0 | 0.0 | 3.2 | 25.1 |
| LoRAstd | 8.1 | 46.7 | 38.8 | 48.4 | 35.5 |
| LoRACoT | 24.6 | 48.3 | 53.0 | 54.3 | 45.1 |
| **Paligemma-2 10B** | | | | | |
| LoRAaug | 38.9 | 18.0 | 24.8 | 16.1 | 24.4 |
| LoRAstd | 9.6 | 39.3 | 39.7 | 58.1 | 36.7 |
| LoRACoT | 8.1 | 61.3 | 43.0 | 61.3 | 43.4 |
## đ Key Findings
- **Chain-of-Thought prompting and fine-tuning** improves both classification and explanation quality.
- **Closed-source models** outperform open-source models in both classification and explanation tasks.
- **Metaphor interpretation** remains challenging for current VLMs, especially in cultural contexts
- **Open-source VLMs** struggle with visual metaphors, with performance dropping further when generating explanations.