Update README.md

README.md

@@ -19,3 +19,45 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+# Muk-Jji-Bba Dataset (SquidGame Series)
+
+**Note**: Please do not use this dataset for training purposes.
+
+## Overview
+
+The "Muk-Jji-Bba" dataset is the first in the SquidGame series, designed to evaluate whether models can understand human behavior. This dataset specifically focuses on the game of Muk-Jji-Bba, a variation of Rock-Paper-Scissors widely played in Korea.
+
+### How Muk-Jji-Bba Works:
+- The attacker tries to match their gesture with the defender’s to win the game.
+- If both players show the same gesture in the next round, the attacker wins.
+- If the attacker’s next gesture loses to the defender’s, the roles switch, and the defender becomes the new attacker.
+- If the attacker’s next gesture beats the defender’s, the attacker keeps their role and the game continues.
+
+The dataset includes 4 rounds per situation. If there is no winner in the final round, the result is a "Tie." The model must choose between Player A (1), Player B (2), or Tie (3).
+
+The labels are evenly distributed across the three possible outcomes.
+
+## Model Performance
+
+| Model               | Accuracy | F1 Score |
+|---------------------|----------|----------|
+| Llama 3.1 (instruct) | 0.8083   | 0.8121   |
+| Llama 3              | 0.775    | 0.7781   |
+| Solar                | 0.7541   | 0.7592   |
+| Qwen                 | 0.7833   | 0.7877   |
+| **Yi-9b-chat (top)**     | 0.8395   | 0.8426   |
+
+## About the Labels
+The labels in the dataset represent the correct outcome for each round:
+- Player A wins (1)
+- Player B wins (2)
+- Tie (3)
+
+The labels are evenly distributed among the three outcomes to ensure balance.
+
+## Stay Tuned
+Look forward to the next series in the SquidGame dataset!
+
+(Evaluation code will be updated soon.)
+