input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 1
value |
|---|---|---|
What is 17 multiplied by 24? | 408 | [paste what model said] |
If all cats are animals and all animals need water do cats need water? | Yes | [paste what model said] |
Can you fit a bus inside a teacup? | No a bus is far too large | [paste what model said] |
Who won the FIFA World Cup in 2022? | Argentina | [paste what model said] |
Translate: میں ٹھیک ہوں | I am fine | [paste what model said] |
What is NOT the capital of Pakistan? | Any city except Islamabad | [paste what model said] |
I have 10 apples give away 3 buy 5 more. How many? | 12 | [paste what model said] |
How many bones are in the adult human body? | 206 | [paste what model said] |
Write a Python function that reverses a string | return s[::-1] | [paste what model said] |
Doctor is to hospital as teacher is to? | School | [paste what model said] |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Blind Spots of Qwen2.5-1.5B
Model Tested
- Model: Qwen/Qwen2.5-1.5B
- Link: https://huggingface.co/Qwen/Qwen2.5-1.5B
- Type: Base language model (not finetuned)
- Parameters: 1.5 Billion
How I Loaded the Model
I used Google Colab with a free T4 GPU.
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_name = "Qwen/Qwen2.5-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,
device_map="auto"
)
Dataset Description
This dataset contains 10 diverse prompts where the model made incorrect or poor predictions. Each row has:
- input: the prompt given to the model
- expected_output: the correct answer
- model_output: what the model actually said
What Fine-Tuning Dataset Would Fix These Errors?
| Error Type | Suggested Dataset |
|---|---|
| Math errors | GSM8K (grade school math problems) |
| Logical reasoning | BIG-Bench, LogiQA |
| Multilingual/Urdu | UrduNLP datasets, mC4 Urdu |
| Factual recall | TriviaQA, Natural Questions |
| Code generation | CodeAlpaca, HumanEval |
How Would I Assemble Such a Dataset?
- Download existing open datasets from Hugging Face
- Manually write diverse Q&A pairs
- Use crowdsourcing to collect varied examples
How Big a Dataset is Needed?
- For targeted fixes: 5,000–20,000 examples is sufficient
- For broad improvement: 100,000+ examples
- Quality matters more than quantity
- Downloads last month
- 5