Update README.md
Browse files
README.md
CHANGED
@@ -6,21 +6,33 @@ license: mit
|
|
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 |
-
|
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 |
-
- `max_tokens=4096`
|
16 |
-
- Inference performed
|
17 |
- Each problem is attempted **128 times**
|
18 |
|
19 |
-
The difficulty score for each problem is computed as:
|
20 |
|
21 |
d_i = 100 × (1 - (# successes / 128))
|
22 |
|
23 |
-
This
|
|
|
|
|
|
|
24 |
|
25 |
-
##
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
- `max_tokens = 4096`
|
16 |
+
- Inference performed using [vLLM](https://github.com/vllm-project/vllm)
|
17 |
- Each problem is attempted **128 times**
|
18 |
|
19 |
+
The difficulty score `d_i` for each problem is computed as:
|
20 |
|
21 |
d_i = 100 × (1 - (# successes / 128))
|
22 |
|
23 |
+
This approach balances the evaluation signal:
|
24 |
+
- A **strong model** would trivially solve easy problems, compressing the difficulty scale.
|
25 |
+
- A **weak model** would fail uniformly, providing poor resolution.
|
26 |
+
- Qwen 2.5-MATH-7B was selected for its **mid-range capabilities**, offering meaningful gradients across a wide spectrum of problems.
|
27 |
|
28 |
+
## Difficulty Estimation on Other Datasets
|
29 |
+
|
30 |
+
We also apply the same difficulty estimation procedure to the following datasets:
|
31 |
+
|
32 |
+
- [Open Reasoner Zero](https://huggingface.co/datasets/lime-nlp/orz_math_difficulty)
|
33 |
+
- [MATH](https://huggingface.co/datasets/lime-nlp/MATH_difficulty)
|
34 |
+
- [GSM8K](https://huggingface.co/datasets/lime-nlp/GSM8K_difficulty)
|
35 |
+
|
36 |
+
## 📬 Contact
|
37 |
+
|
38 |
+
For questions or feedback, feel free to reach out to [**Taiwei Shi**](https://maksimstw.github.io/) at [[email protected]](mailto:[email protected]).
|