Upload README.md with huggingface_hub

README.md

@@ -12,5 +12,5 @@ tags:
 
 Curriculum-filtered subset of [OctoReasoner/verl_codeql1_aggressive](https://huggingface.co/datasets/OctoReasoner/verl_codeql1_aggressive).
 
-Kept rows satisfy **weak <= 5/8 AND strong >= 2/8**, where weak/strong = number of correct rollouts (of 8, strict \boxed{} scoring) from Qwen2.5-Coder-1.5B-Instruct and -32B-Instruct.
-5,788 of 14,432 rows kept. Pure verl schema.
+**CoT variant.** Models were prompted to *think first, then box* (max_tokens=4096). Kept rows satisfy **weak <= 5/8 AND strong >= 2/8**, where weak/strong = number of correct rollouts (of 8, strict \boxed{} scoring) from Qwen2.5-Coder-1.5B/-32B-Instruct.
+5591 rows kept. Pure verl schema.