Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,22 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
---
|
4 |
+
|
5 |
+
## Difficulty Estimation on Open Reasoner Zero
|
6 |
+
We annotate the entire [**Open Reasoner Zero**]((https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-7B)) dataset with a **difficulty score** based on the performance of the [Qwen 2.5-MATH-7B](https://huggingface.co/Qwen/Qwen2.5-Math-7B) model. This provides an adaptive signal for curriculum construction.
|
7 |
+
Open Reasoner Zero is a curated a dataset of 57,000 reasoning-intensive problems used to train and evaluate reinforcement learning-based methods for large language models.
|
8 |
+
|
9 |
+
### Difficulty Scoring Method
|
10 |
+
|
11 |
+
Difficulty scores are estimated using the Qwen 2.5-MATH-7B model with the following generation settings:
|
12 |
+
|
13 |
+
- `temperature = 0.6`
|
14 |
+
- `top_p = 0.9`
|
15 |
+
- Inference performed via [vLLM](https://github.com/vllm-project/vllm)
|
16 |
+
- Each problem is attempted **128 times**
|
17 |
+
|
18 |
+
The difficulty score for each problem is computed as:
|
19 |
+
|
20 |
+
d_i = 100 × (1 - (# successes / 128))
|
21 |
+
|
22 |
+
This scoring approach ensures a balanced estimation: a strong model would trivially succeed on all problems, undermining difficulty measurement, while a weak model would fail uniformly, limiting the usefulness of the signal. Qwen 2.5-MATH-7B was chosen for its **mid-range capabilities**, providing **informative gradients** in problem difficulty across the dataset.
|