Commit
·
0b36220
1
Parent(s):
b03a5bc
Description for the dataset.
Browse files
README.md
CHANGED
|
@@ -1,3 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
dataset_info:
|
| 3 |
- config_name: do-not-respond
|
|
|
|
| 1 |
+
# SRBench
|
| 2 |
+
|
| 3 |
+
This dataset is a collection of six different types of questions that are observed to be difficult for reasoning models to answer but easy for humans.
|
| 4 |
+
|
| 5 |
+
We have observed that models like DeepSeek R1 and its derivatives struggle to follow certain instructions and tend to focus on the wrong parts of the questions. As a result, they spend a significant amount of time during the thinking process and sometimes fail to answer the questions correctly.
|
| 6 |
+
|
| 7 |
---
|
| 8 |
dataset_info:
|
| 9 |
- config_name: do-not-respond
|