adding additional reference and my current expereiments

README.md

@@ -27,7 +27,7 @@ datasets:
 8. [Optimization Techniques](#8-optimization-techniques)
 9. [Lessons Learned](#9-lessons-learned-from-implementing-muon)
 10. [Concludsion](#10-conclusion)
-10. [Extended Work](#11-extended-work)
+11. [Extended Work](#11-extended-work)
 
 ## 🧪 Try It Yourself
 
@@ -128,6 +128,8 @@ This hybrid approach gets the best of both worlds: Muon's efficient matrix updat
 ### Training Stages
 - ✅ Pre-training from scratch
 - ✅ Domain adaptation
+- 🔬?RL training for math reasoning tasks 
+- 👉🏻 I'm currently running experiements to investigate the training dynamicas, stay turned for update!
 - ❌ Fine-tuning (low-rank updates like LoRA are preferable)
 - ❌ Alignment stages (RLHF/DPO)
 
@@ -256,9 +258,12 @@ As you build and train your own models, consider Muon for hidden layer optimizat
 
 
 ## 11. Extended Work
-For the distributed (DP × TP) implementation built for CPU/Gloo environments, see:
+- For the distributed (DP × TP) implementation built for CPU/Gloo environments, see:
 
-[🧩 The "Muon is Scalable" Blueprint: A Distributed Muon Engineering Breakdown (CPU-Friendly, Tutorial Style)](https://huggingface.co/datasets/bird-of-paradise/muon-distributed)
+ [🧩 The "Muon is Scalable" Blueprint: A Distributed Muon Engineering Breakdown (CPU-Friendly, Tutorial Style)](https://huggingface.co/datasets/bird-of-paradise/muon-distributed)
+
+- For those if you who are intereted in validation work on Moonshot's "Muon is scalable for LLM trianing", 👉🏻 check out
+  [🔬 Distributed Muon: Field Notes & Reproducibility Artifacts](https://huggingface.co/datasets/bird-of-paradise/muon-distributed-reproducibility)
 
 ---