Update README.md

README.md

@@ -1,8 +1,31 @@
+---
+license: mit
+task_categories:
+- text-generation
+- token-classification
+pretty_name: Terminal Recording Group Boundary Timestamps
+size_categories:
+- n < 1K
+tags:
+- terminal
+- xml
+- boundary-detection
+---
+
+# Terminal Recording Group Boundary Dataset
+
+This dataset consists of raw terminal recording logs in XML format, paired with ground-truth timestamps for shell prompt boundaries.
+
+## Training Approach: Event Chunking
+
+Terminal recordings are essentially long, continuous streams of text data. To make this data useful for downstream tasks, we must first break this stream into logical "events" (e.g., a single command execution and its resulting output). 
+**Model 0** is specifically designed for this segmentation task. Instead of just looking for text patterns, the model is trained to identify the precise temporal boundary where one event ends and the next begins.
+
 ## Generation Logic & Sliding Windows (Important)
 
-The 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.
+The label the model must predict is the **boundary timestamp**—the exact moment the shell prompt returns, signaling the end of an event.
 
-To generate training data, we use a sliding window approach encompassing exactly **three consecutive boundary timestamps** (**T1**, **T2**, and **T3**) for each input:
+To generate robust training data, we use a sliding window approach encompassing exactly **three consecutive boundary timestamps** (**T1**, **T2**, and **T3**) for each input:
 
 * **T1 (The Anchor):** The known starting timestamp of the current event.
 * **T2 (The Target):** The actual event boundary the model must predict.
@@ -10,12 +33,16 @@ To generate training data, we use a sliding window approach encompassing exactly
 
 ### Why include the third timestamp (T3)?
 
-Each input is fed to the model containing all `<user_input>` and `<system_output>` XML tags from **T1** all the way to **T3**. If the input simply stopped at **T2**, 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 **T3**, the model is forced to semantically identify the actual boundary dividing the two separate events within the continuous log.
+Each training input contains all `<user_input>` and `<system_output>` XML tags from **T1** all the way to **T3**. 
+
+If the input window stopped  at **T2**, the task would be trivial: the model would eventually realize it just needs to extract the very last timestamp present in the provided text. By extending the context to **T3**, we include a "distractor" event. This forces the model to semantically understand the terminal log and identify the **internal** boundary dividing the two separate events, rather than just pointing to the end of the file.
 
 ### 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).
+
+1. **The Input Window:** Contains all XML chunks from `0.007194` to `37.178971`.
+2. **The Content:** This window actually covers two distinct events:
+    * **Event A:** Context between `0.007194` and `21.987222`.
+    * **Event B:** Context between `21.987222` and `37.178971`.
+3. **The Target Output:** `21.987222` (the dividing boundary).