input stringclasses 10
values | model_output stringclasses 5
values | expected_output stringclasses 9
values |
|---|---|---|
44 * 72 | 2768 | 3168 |
0.25 + 0.33 + 0.42 | 0.999999999999999 | 1.0 |
Add 0.25*12 + 0.33*12 + 0.42*12 | incomplete | 12 |
Each apple costs 1.22 and I buy 9. Total price? | stuck thinking | 10.98 |
Pizza cut into 8 slices, eat 3. Fraction left? | stuck thinking | 5/8 |
Simplify 18/24 | stuck thinking | 3/4 |
How many months have 28 days? | stuck thinking | 12 |
Remember number 8472. What number? | could not recall | 8472 |
A father is 40 years old, son 10. In how many years will father be twice as old? | stuck thinking | 20 |
Water boils at 100°C at sea level. Will it boil faster or slower on top of a mountain? | stuck thinking | faster |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Nanbeige Blind Spots Dataset
This dataset contains 10 data points where the Nanbeige4-3B-Base model makes mistakes or gets stuck thinking. It includes the input prompt, the model's output, and the expected output.
Model Tested: Nanbeige4-3B-Base
Dataset Fields
input: The input prompt given to the model.model_output: What the model returned.expected_output: The correct answer.
Example Data Points
| input | model_output | expected_output |
|---|---|---|
| 44 * 72 | 2768 | 3168 |
| 0.25 + 0.33 + 0.42 | 0.999999999999999 | 1.0 |
| Pizza cut into 8 slices, eat 3 | stuck thinking | 5/8 |
| ... | ... | ... |
Suggested Fine-Tuning
To improve the model, it could be fine-tuned on basic arithmetic, fractions, percentages, memory tasks, and real-world reasoning problems.
A dataset for fine-tuning could be assembled by combining:
- Mathematical problem datasets (arithmetic, fractions, percentages)
- Short reasoning tasks (age, speed, basic physics, everyday questions)
- Memory/recall tasks
The dataset does not need to be huge; a few thousand well-curated examples (3k–5k) should start improving the model on these blind spots.
- Downloads last month
- 6