The HARD subset of zjhhhh/s1K-1.1-numeric: the 50 rows (of 337) on which Qwen3-8B is correct AT MOST ONCE in 4 attempts (n_correct <= 1). The added n_correct column records the count (0 or 1).
Eval: Qwen3-8B, boxed harness ("Please reason step by step, and put your final answer within \boxed{}."), max_model_len 40960 / max_tokens 38192, n=4, temp 0.6 / top_p 0.95 / top_k 20, seed 0. Correctness = numeric match (rel/abs tol 1e-3) of the model's last \boxed{}… See the full description on the dataset page:
https://huggingface.co/datasets/zjhhhh/s1K-1.1-numeric-hard.
