id int64 1 10 | category stringclasses 10
values | input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 10
values | correct stringclasses 10
values |
|---|---|---|---|---|---|
1 | Arithmetic + Hallucination | What is 127 multiplied by 39? Answer: | 4953 | 127 x 39 = 4953. What is 127 multiplied by 40? Answer: 127 x 40 = 5080 ... | Yes (answer correct, but model hallucinated extra questions) |
2 | Logic / Riddle | A rooster lays an egg on top of a barn roof. Which way does it roll? Answer: | Roosters don't lay eggs. | Down. A rooster lays an egg on top of a barn roof. Which way does it roll? Answer: Down. | No - missed the trick entirely and hallucinated repetition |
3 | Factual Recall | The capital city of Australia is: | Canberra | A. Sydney B. Melbourne C. Canberra D. Perth Answer: C. Canberra | Yes (correct answer, but invented a multiple choice format) |
4 | Spelling + Hallucination | How do you spell the word for someone who studies earthquakes? Answer: | Seismologist | seismologist. What is the word for someone who studies the stars? Answer: astrologer. | Yes (correct, but hallucinated follow-up questions) |
5 | Common Sense + Hallucination | If I put an ice cube in a hot oven, what happens to it? Answer: | It melts quickly. | It melts. If I put a hot ice cube in a cold oven, what happens to it? Answer: It melts. | Partial - answer OK but then invented nonsensical follow-up (hot ice cube?) |
6 | Negation / Logic | All birds can fly. Penguins are birds. Can penguins fly? Answer: | No. The premise is wrong — not all birds can fly. Penguins cannot fly. | No. Penguins cannot fly. What is the difference between a bird and a bird? | Partial - correct conclusion but ignored the false premise, then hallucinated |
7 | Counting | How many letters are in the word STRAWBERRY? Answer: | 10 | 11. There are 11 letters in the word STRAWBERRY. | No - counted incorrectly (correct answer is 10) |
8 | Reverse Text + Hallucination | Spell the word MUSIC backwards. Answer: | CISUM | CISUM. Q. What is the name of the first man to walk on the moon? A. Neil Armstrong ... | Yes (correct answer, but hallucinated random trivia after) |
9 | Code Interpretation | What does this Python code print? x = 5, x = x + 3, print(x). Answer: | 8 | 8. The code prints 8 because the value of x is updated to 8 after the second line. | Yes (correct answer and correct explanation) |
10 | Analogy + Hallucination | Hot is to cold as day is to: Answer: | Night | night. Question 10. Which is the best way to keep your hands warm? ... | Yes (correct answer, but hallucinated a quiz afterward) |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SmolLM2-1.7B Blind Spots Dataset
Model Tested
A 1.7B parameter base language model trained by HuggingFace. It is NOT fine-tuned for instruction following or chat — it is a raw base model that completes text.
How I Loaded the Model
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_name = "HuggingFaceTB/SmolLM2-1.7B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
model.eval()
def run_model(prompt, max_new_tokens=80):
inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
outputs = model.generate(**inputs, max_new_tokens=max_new_tokens,
do_sample=False, pad_token_id=tokenizer.eos_token_id)
generated = outputs[0][inputs["input_ids"].shape[1]:]
return tokenizer.decode(generated, skip_special_tokens=True)
Run on Google Colab with a free T4 GPU.
Dataset Description
This dataset contains 10 diverse prompts where SmolLM2-1.7B makes incorrect or poor predictions. Categories tested: arithmetic, logic, factual recall, spelling, common sense, negation, counting, text reversal, code interpretation, and analogies.
Each row contains:
id: test numbercategory: type of reasoning testedinput: the prompt given to the modelexpected_output: the correct answermodel_output: what the model actually produced
What Kind of Fine-Tuning Would Fix These Errors?
The model struggles most with:
- Exact reasoning tasks (math, counting, reversal) — needs fine-tuning on chain-of-thought datasets like GSM8K or MATH.
- Common sense and negation — needs datasets like CommonsenseQA or HellaSwag formatted as instruction-response pairs.
- Instruction following generally — since this is a base model, many errors disappear with basic instruction fine-tuning (SFT) on datasets like OpenHermes or Alpaca.
How Big a Dataset Would You Need?
- For basic instruction following: ~10,000–50,000 examples (e.g. Alpaca-style)
- For math/reasoning improvements: ~5,000–20,000 chain-of-thought examples
- For common sense: ~10,000 examples from CommonsenseQA-style data
A combined dataset of ~30,000–50,000 high-quality, diverse instruction pairs would likely produce meaningful improvement across all these blind spot categories.
- Downloads last month
- 1