nirmalendu01/animal_dataset_activations

Animal Dataset Activations

This dataset contains neural network activations captured from various models processing the animal_dataset.

Dataset Structure

The dataset is organized by model (subset) and layer (split):

• Model (subset): Each model has its own directory
• Layer (split): Each layer within a model has its own directory with parquet files

Files are organized as: {model_name}/{layer_name}/*.parquet

Loading the Dataset

The dataset uses HuggingFace's subset/split API where:

• Repository: nirmalendu01/animal_dataset_activations
• Model (subset): Use name="model_name" parameter
• Layer (split): Use split="layer_name" parameter

Basic Usage

from datasets import load_dataset

# Format: load_dataset(repo_name, name="model_name", split="layer_name")
ds = load_dataset('nirmalendu01/animal_dataset_activations', name='model_name', split='layer_name')

Examples

Qwen_Qwen2.5-7B-Instruct

from datasets import load_dataset

# Load layer 'model_layers_15' from Qwen_Qwen2.5-7B-Instruct
ds = load_dataset('nirmalendu01/animal_dataset_activations', 
    name='Qwen_Qwen2.5-7B-Instruct', 
    split='model_layers_15')

# Available layers for Qwen_Qwen2.5-7B-Instruct: model_layers_15, model_layers_16, model_layers_17, model_layers_18, model_layers_19...

clip

from datasets import load_dataset

# Load layer 'image_visual_transformer_resblocks_12' from clip
ds = load_dataset('nirmalendu01/animal_dataset_activations', 
    name='clip', 
    split='image_visual_transformer_resblocks_12')

# Available layers for clip: image_visual_transformer_resblocks_12, image_visual_transformer_resblocks_13, image_visual_transformer_resblocks_14, image_visual_transformer_resblocks_15, image_visual_transformer_resblocks_16...

List Available Models and Layers

from huggingface_hub import list_repo_files

# List all files to see available models and layers
files = list_repo_files(repo_id='nirmalendu01/animal_dataset_activations', repo_type='dataset')

# Files are organized as: {{model_name_sanitized}}/{{layer_name}}/*.parquet
# Note: Model names with "/" are sanitized to "_" in folder names
#       (e.g., "facebook/dinov2-base" -> "facebook_dinov2-base")

Alternative: Direct Parquet Loading

If you prefer to load directly from parquet files (without the dataset script):

from datasets import load_dataset

# Note: Use sanitized model name (replace "/" with "_")
model_name_sanitized = "facebook_dinov2-base"  # for "facebook/dinov2-base"
layer_name = "encoder_layer_11"

ds = load_dataset('parquet', 
    data_files=f'https://huggingface.co/datasets/nirmalendu01/animal_dataset_activations/resolve/main/{{model_name_sanitized}}/{{layer_name}}/*.parquet')

Note: The subset/split API (using name and split parameters) is preferred as it's more convenient and handles model name sanitization automatically.

Load a specific layer:

from datasets import load_dataset

# Load a specific layer
ds = load_dataset('parquet', 
    data_files=f'https://huggingface.co/datasets/nirmalendu01/animal_dataset_activations/resolve/main/{model_name}/{layer_name}/*.parquet')

Load all layers for a model:

# List all layers first, then load each
from huggingface_hub import list_repo_files
files = list_repo_files(repo_id='nirmalendu01/animal_dataset_activations', repo_type='dataset')
# Filter and load as needed

Layer Statistics

CLIP

Layer Name	Samples	Files	Total Size (MB)	Avg File Size (MB)	Activation Shape
image_visual_transformer_resblocks_12	1,000	10	1550.16	155.02	(257, 1024)
image_visual_transformer_resblocks_13	1,000	10	1550.56	155.06	(257, 1024)
image_visual_transformer_resblocks_14	1,000	10	1551.98	155.20	(257, 1024)
image_visual_transformer_resblocks_15	1,000	10	1552.21	155.22	(257, 1024)
image_visual_transformer_resblocks_16	1,000	10	1552.27	155.23	(257, 1024)
image_visual_transformer_resblocks_17	1,000	10	1553.21	155.32	(257, 1024)
image_visual_transformer_resblocks_18	1,000	10	1554.12	155.41	(257, 1024)
image_visual_transformer_resblocks_19	1,000	10	1554.53	155.45	(257, 1024)
image_visual_transformer_resblocks_20	1,000	10	1553.63	155.36	(257, 1024)
image_visual_transformer_resblocks_21	1,000	10	1552.98	155.30	(257, 1024)
image_visual_transformer_resblocks_22	1,000	10	1553.81	155.38	(257, 1024)
image_visual_transformer_resblocks_23	1,000	10	1552.81	155.28	(257, 1024)
text_transformer_resblocks_10	1,000	10	348.22	34.82	(77, 768)
text_transformer_resblocks_11	1,000	10	348.06	34.81	(77, 768)
text_transformer_resblocks_6	1,000	10	348.55	34.85	(77, 768)
text_transformer_resblocks_7	1,000	10	348.58	34.86	(77, 768)
text_transformer_resblocks_8	1,000	10	313.89	31.39	(77, 768)
text_transformer_resblocks_9	1,000	10	348.73	34.87	(77, 768)

facebook/dinov2-base

Layer Name	Samples	Files	Total Size (MB)	Avg File Size (MB)	Activation Shape
encoder_layer_10	1,000	10	1161.21	116.12	(257, 768)
encoder_layer_11	1,000	10	1161.13	116.11	(257, 768)
encoder_layer_6	1,000	10	1161.34	116.13	(257, 768)
encoder_layer_7	1,000	10	1160.79	116.08	(257, 768)
encoder_layer_8	1,000	10	1160.67	116.07	(257, 768)
encoder_layer_9	1,000	10	1160.83	116.08	(257, 768)

Qwen/Qwen2.5-7B-Instruct

Layer Name	Samples	Files	Total Size (MB)	Avg File Size (MB)	Activation Shape
model_layers_15	1,000	10	93.42	9.34	(13, 3584)
model_layers_16	1,000	10	93.42	9.34	(13, 3584)
model_layers_17	1,000	10	93.38	9.34	(13, 3584)
model_layers_18	1,000	10	93.38	9.34	(13, 3584)
model_layers_19	1,000	10	93.37	9.34	(13, 3584)
model_layers_20	1,000	10	93.35	9.33	(13, 3584)
model_layers_21	1,000	10	93.30	9.33	(13, 3584)
model_layers_22	1,000	10	93.24	9.32	(13, 3584)
model_layers_23	1,000	10	93.20	9.32	(13, 3584)
model_layers_24	1,000	10	93.17	9.32	(13, 3584)
model_layers_25	1,000	10	93.05	9.30	(13, 3584)
model_layers_26	1,000	10	93.02	9.30	(13, 3584)
model_layers_27	1,000	10	93.25	9.32	(13, 3584)

stabilityai/stable-diffusion-xl-base-1.0

Layer Name	Samples	Files	Total Size (MB)	Avg File Size (MB)	Activation Shape
unet_output_step_0_cond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_0_uncond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_1_cond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_1_uncond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_2_cond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_2_uncond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_3_cond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_3_uncond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_4_cond	1,000	10	118.34	11.83	(4, 128, 128)
unet_output_step_4_uncond	1,000	10	118.34	11.83	(4, 128, 128)

llava-hf/llava-onevision-qwen2-7b-ov-hf

Layer Name	Samples	Files	Total Size (MB)	Avg File Size (MB)	Activation Shape
llm_layers_16	1,000	20	21047.47	1052.37	(2949, 3584)
llm_layers_17	1,000	20	20921.14	1046.06	(2949, 3584)
llm_layers_18	1,000	20	20921.11	1046.06	(2949, 3584)
llm_layers_19	1,000	20	20921.04	1046.05	(2949, 3584)
llm_layers_20	1,000	20	20920.97	1046.05	(2949, 3584)
llm_layers_21	1,000	20	20920.88	1046.04	(2949, 3584)
llm_layers_22	950	19	19874.71	1046.04	(2949, 3584)
llm_layers_23	1,000	20	20920.71	1046.04	(2949, 3584)
llm_layers_24	1,000	20	20920.69	1046.03	(2949, 3584)
llm_layers_25	1,000	20	20920.59	1046.03	(2949, 3584)
llm_layers_26	1,000	20	20920.32	1046.02	(2949, 3584)
llm_layers_27	1,000	20	20921.16	1046.06	(2949, 3584)
vision_layers_14	1,000	10	1511.21	151.12	(729, 1152)
vision_layers_15	1,000	10	1511.37	151.14	(729, 1152)
vision_layers_16	1,000	10	1511.30	151.13	(729, 1152)
vision_layers_17	1,000	10	1511.20	151.12	(729, 1152)
vision_layers_18	900	9	1360.24	151.14	(729, 1152)
vision_layers_19	1,000	10	1511.43	151.14	(729, 1152)
vision_layers_20	1,000	10	1511.35	151.14	(729, 1152)
vision_layers_21	1,000	10	1511.23	151.12	(729, 1152)
vision_layers_22	1,000	10	1511.19	151.12	(729, 1152)
vision_layers_23	1,000	10	1511.19	151.12	(729, 1152)
vision_layers_24	1,000	10	1511.18	151.12	(729, 1152)
vision_layers_25	1,000	10	1511.24	151.12	(729, 1152)

• llava-hf/llava-onevision-qwen2-7b-ov-hf: Neural network activations
• stabilityai/stable-diffusion-xl-base-1.0: Stable Diffusion XL - Text encoder and UNet activations
• Qwen/Qwen2.5-7B-Instruct: Qwen2.5-7B-Instruct - Large Language Model activations
• facebook/dinov2-base: DINOv2 - Vision Transformer activations
• CLIP: CLIP (Contrastive Language-Image Pre-training) - Vision and text encoder activations

Source Dataset

Activations were captured from: nirmalendu01/animal_dataset

Format

Each parquet file contains:

• sample_key: Unique identifier for the sample
• activation_shape: Shape of the activation tensor
• activation_data: Flattened activation data (can be reshaped using activation_shape)
• Additional metadata fields (e.g., caption, prompt) depending on the model

Usage Example

from datasets import load_dataset
import torch
import numpy as np

# Load activations for a specific layer
ds = load_dataset('parquet', 
    data_files=f'https://huggingface.co/datasets/nirmalendu01/animal_dataset_activations/resolve/main/clip/text_transformer_resblocks_0/*.parquet')

# Access a sample
sample = ds['train'][0]

# Reconstruct the activation tensor
shape = tuple(sample['activation_shape'])
activation = torch.from_numpy(np.array(sample['activation_data']).reshape(shape))

print(f"Activation shape: {activation.shape}")
print(f"Metadata: {sample.get('caption', 'N/A')}")

Notes

• Parquet files are chunked (100 samples per file by default) to support streaming
• Activation tensors are stored as flattened arrays and can be reshaped using the activation_shape field
• All activations are captured on the animal dataset with 1000 samples (subset of full dataset)

Citation

If you use this dataset, please cite the original animal dataset and the models used.