input stringclasses 10
values | expected_output stringclasses 9
values | model_output stringclasses 9
values |
|---|---|---|
How many months have 28 days? | 12 | 4 |
Reverse the word 'drawer' | reward | rewrda |
How many letters are in the word KNOWLEDGE? | 9 | 5 |
How many 'p' letters are in pineapple? | 3 | 4 |
Reverse the word 'algorithm' | mhtirogla | htiroglaM |
How many letters are in SUBDERMATOGLYPHIC? | 17 | 12 |
The father of Mary has five daughters: Nana, Nene, Nini, Nono. What is the fifth daughter? | Mary | Nene |
What is (37 × 24) + 58? | 946 | 1066 |
How many animals of each species did Moses take on the Ark? | None (it was Noah) | Two animals |
If two’s company and three’s a crowd, what are four and five? | 9 | A pair |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Model Blind Spots Dataset
Model Tested
Model: microsoft/phi-2
Link: https://huggingface.co/microsoft/phi-2
How the Model Was Loaded
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "microsoft/phi-2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
## Description
This dataset contains prompts where the model produced incorrect outputs.
The blind spots include:
- Letter counting errors
- Word reversal mistakes
- Arithmetic mistakes
- Logic puzzles
- Knowledge confusion
## How to Fix These Errors
The model could be improved by fine-tuning on datasets containing:
- arithmetic reasoning tasks
- string manipulation tasks
- logic puzzles
## Possible datasets for fine-tuning
- GSM8K
- BIG-Bench
- synthetic reasoning datasets
## Estimated Dataset Size
An estimated dataset size of **10k–50k examples** may be required to significantly reduce these errors.
- Downloads last month
- 1