Upload README.md with huggingface_hub

README.md

@@ -24,6 +24,13 @@ This repo hosts the **eval reports and ladder plots** from the full
 Qwen3.8-27B study: three SFT arms (incl. a 4x-volume collapse probe), one
 GRPO/LadderRL arm, and the ThinkingCap external control.
 
+## Training data
+
+Arm training combined self-generated, correctness-filtered incident lessons
+with a constant grounding layer of 5,000 rows from the Hugging Face dataset
+`stindardlogic/devops-kubernetes-sft-100k` (100k rows, Apache-2.0),
+stratified toward troubleshooting and debugging categories.
+
 ## Limitations (read before citing)
 
 - **68-probe eval set** — one item moves accuracy by ~1.5%; checkpoint