Upload README.md with huggingface_hub

README.md

@@ -30,6 +30,35 @@ this dataset contains pre-rendered word stimulus images used to evaluate how wel
 
 each image is a 224x224 black-background PNG with white text, rendered in arial at size 22, centred.
 
+### example stimuli for the target word "design"
+
+<table>
+  <tr>
+    <td align="center"><b>ID</b><br>(identity)</td>
+    <td align="center"><b>TL12</b><br>(transposed 1-2)</td>
+    <td align="center"><b>DL-1M</b><br>(deleted middle)</td>
+    <td align="center"><b>SN-M</b><br>(substituted middle)</td>
+    <td align="center"><b>RF</b><br>(reversed full)</td>
+    <td align="center"><b>ALD-ARB</b><br>(all different)</td>
+  </tr>
+  <tr>
+    <td align="center"><img src="prime_data/design/ID.png" width="120"></td>
+    <td align="center"><img src="prime_data/design/TL12.png" width="120"></td>
+    <td align="center"><img src="prime_data/design/DL-1M.png" width="120"></td>
+    <td align="center"><img src="prime_data/design/SN-M.png" width="120"></td>
+    <td align="center"><img src="prime_data/design/RF.png" width="120"></td>
+    <td align="center"><img src="prime_data/design/ALD-ARB.png" width="120"></td>
+  </tr>
+  <tr>
+    <td align="center">DESIGN</td>
+    <td align="center">EDSIGN</td>
+    <td align="center">DSIGN</td>
+    <td align="center">DESIHN</td>
+    <td align="center">NGISE</td>
+    <td align="center">CBHAUX</td>
+  </tr>
+</table>
+
 ### what's included
 
 ```
@@ -47,7 +76,7 @@ metadata/
 
 | code | description |
 |------|-------------|
-| ID | identity (e.g., design -> design) |
+| ID | identity (e.g., prime and target are both "design") |
 | TL12 | transposed letters positions 1-2 |
 | TL-I | transposed letters internal |
 | TL56 | transposed letters positions 5-6 |
@@ -101,7 +130,7 @@ img = Image.open(prime_dir / target / f"{condition}.png")
 
 the core analysis computes kendall's tau between model cosine-similarity patterns and human priming scores across the 28 conditions. see the [source code repository](https://github.com/Don-Yin/Orthographic-DNN) for the full pipeline:
 
-1. fine-tune pretrained torchvision models on word classification (training images not included here - generate with `generate_data.py`)
+1. fine-tune pretrained torchvision models on word classification (training images not included here; generate with `generate_data.py`)
 2. extract layer-wise activations for each prime image pair (identity vs. condition)
 3. compute cosine similarity at each layer
 4. correlate with human priming scores using kendall's tau
@@ -118,7 +147,7 @@ this requires the font files (not redistributable) and generates images with con
 
 ## models evaluated
 
-alexnet, densenet169, efficientnet-b1, resnet50, resnet101, vgg16, vgg19, vit-b/16, vit-b/32, vit-l/16, vit-l/32 - all initialised from imagenet pretrained weights via torchvision.
+alexnet, densenet169, efficientnet-b1, resnet50, resnet101, vgg16, vgg19, vit-b/16, vit-b/32, vit-l/16, vit-l/32, all initialised from imagenet pretrained weights via torchvision.
 
 ## citation