Add training dataset link to model card

README.md

@@ -12,6 +12,8 @@ tags:
   - twenty-questions
 language:
   - en
+datasets:
+  - david-ar/20q-dataset
 ---
 
 # TwentyQ — The World's Smallest Chat Model
@@ -76,6 +78,10 @@ while True:
 - **Regular questions**: `Yes`, `No`, `Probably`, `Doubtful`, `Maybe`, `Unknown`
 - **Guesses**: `Yes`, `No`, `Close`
 
+## Training Data
+
+Trained on [`david-ar/20q-dataset`](https://huggingface.co/datasets/david-ar/20q-dataset), a corpus of 9,600 Twenty Questions conversations covering 1,200 objects across 156 features. Answers include graded confidence levels (Yes, No, Probably, Doubtful) rather than binary labels, giving the model finer-grained signal for learning association strengths.
+
 ## How It Works
 
 The model is a weight matrix mapping 156 features (questions) to 1,200 output classes (objects). Each weight is 2 bits encoding polarity and strength. Inference is a scored lookup — no matrix multiplication, no attention, no backprop. Just XOR and addition.