Refine dataset artwork and pipeline presentation

README.md

@@ -87,11 +87,7 @@ configs:
 </div>
 
 <p align="center">
-  <img src="assets/data.png" alt="Bactrainus HotpotQA training-data artwork" width="380">
-</p>
-
-<p align="center">
-  <img src="architecture.svg" alt="Bactrainus modular architecture" width="960">
+  <img src="assets/data.png" alt="Bactrainus HotpotQA training-data artwork" width="507">
 </p>
 
 ## At a glance
@@ -109,6 +105,16 @@ configs:
 | Storage | Sharded Parquet with Zstandard compression |
 | License | CC BY-SA 4.0 |
 
+### How the data follows the Bactrainus pipeline
+
+The eight configurations expose the same evidence flow in progressively more specialized training views. The structured records preserve the question, all supplied candidate paragraphs, answer, and supporting facts. The SFT views then supervise paragraph selection, optional question decomposition, sentence-level evidence selection, and answer generation, either as separate modules or as a joint selector-reader task. Every view retains the same `source_id`, so examples can be joined without relying on row position.
+
+<p align="center">
+  <img src="architecture.svg" alt="Bactrainus modular architecture" width="900">
+</p>
+
+<p align="center"><em>Canonical records are transformed into task-specific conversations while preserving a shared source identity across the modular pipeline.</em></p>
+
 ## Load the data
 
 ```python