Update README.md

README.md

@@ -5,7 +5,11 @@ task_categories:
 - token-classification
 pretty_name: Terminal Recording Group Boundary Timestamps
 size_categories:
-- 10K<n<100K
+- n < 1K
+tags:
+- terminal
+- xml
+- boundary-detection
 ---
 
 # Terminal Recording Group Boundary Dataset
@@ -15,12 +19,33 @@ This dataset consists of raw terminal recording logs in XML format, paired with
 ## Dataset Structure
 
 Each entry in the `.jsonl` file contains:
-- **instruction**: A specialized prompt identifying the `start_timestamp` and the extraction rules.
-- **input**: Raw XML chunks including `<user_input>` and `<system_output>` tags, preserving all ANSI escape codes.
-- **output**: The target timestamp of the **first-most** shell prompt found after the starting point.
+* **instruction**: A specialized prompt identifying the `start_timestamp` and the extraction rules.
+* **input**: Raw XML chunks including `<user_input>` and `<system_output>` tags, preserving all ANSI escape codes.
+* **output**: The target timestamp of the **first-most** shell prompt found after the starting point.
 
 ## Purpose
+
 This data is designed to train **Model 0** to navigate messy terminal logs and precisely identify the point where a command execution ends and the shell prompt returns.
 
-## Generation
-Generated using a sliding window script that captures sequences between established boundary markers.
+## Generation Logic & Sliding Windows (Important)
+
+The ultimate goal of **Model 0** is to accurately chunk continuous terminal recordings into distinct "events." The label the model must predict is the **boundary timestamp**—the exact moment one event ends and a new one begins.
+
+To generate robust training data, we use a sliding window approach encompassing exactly **three consecutive boundary timestamps** ($T_1$, $T_2$, and $T_3$) for each input:
+
+* **$T_1$ (The Anchor):** The known starting timestamp of the current event.
+* **$T_2$ (The Target):** The actual event boundary the model must predict.
+* **$T_3$ (The Context Cutoff):** The end of the input window. 
+
+
+### Why include the third timestamp ($T_3$)?
+
+Each input is fed to the model containing all `<user_input>` and `<system_output>` XML tags from $T_1$ all the way to $T_3$. If the input simply stopped at $T_2$, the task would be trivial: the model would just learn a shortcut to extract the very last timestamp present in the text. By including the data up to $T_3$, the model is forced to semantically identify the actual boundary dividing the two separate events within the continuous log.
+
+### Example
+
+Assume a recording has three sequential boundaries: `0.007194`, `21.987222`, and `37.178971`.
+* **Input:** Contains all XML chunks from `0.007194` to `37.178971`.
+* **Event 1:** The context between `0.007194` and `21.987222`.
+* **Event 2:** The context between `21.987222` and `37.178971`.
+* **Target Output:** `21.987222` (the dividing boundary between Event 1 and Event 2).