input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 10
values | error_type stringclasses 10
values |
|---|---|---|---|
What is 17 x 24? | 408 | Model gives wrong number | Arithmetic error |
What comes after Wednesday? | Thursday | Model skips or misnumbers days | Day ordering |
Translate Good morning to Yoruba | E kaaro | Model outputs anglicized or wrong text | Low-resource language |
Write a Python function to reverse a string | def reverse(s): return s[::-1] | Model produces broken logic | Code generation error |
Who wrote Things Fall Apart? | Chinua Achebe | Model hallucinates a different author | Knowledge gap |
Is 97 a prime number? | Yes 97 is prime | Model says no or gives wrong justification | Logical reasoning |
What is the opposite of generous? | Stingy or miserly | Model gives weakly related antonym | Semantic understanding |
List 3 African countries bordering the Atlantic | Nigeria, Ghana, Senegal | Model lists landlocked or wrong countries | Geographic knowledge |
The sky is blue because | Rayleigh scattering of sunlight | Model gives vague or wrong explanation | Scientific reasoning |
I have 5 apples give away 3 buy 4 more how many? | 6 | Model miscounts or skips a step | Multi-step arithmetic |
Model Tested
HuggingFaceTB/SmolLM2-1.7B
How I loaded it
Used HuggingFace Transformers with AutoModelForCausalLM on Google Colab (T4 GPU, float16). Greedy decoding (do_sample=False) for reproducibility. View Colab Notebook
Blind Spots Found
The model struggled with: multi-step arithmetic, low-resource languages (Yoruba), African geographic knowledge, code generation, and logical reasoning.
Fine-tuning Dataset Recommendation
- GSM8K / MATH for arithmetic reasoning
- FLORES-200 for low-resource language translation
- AfriQA for African knowledge coverage
- HumanEval for code correctness
A dataset of ~50,000–100,000 diverse examples with chain-of-thought annotations would be sufficient to address these failure modes.
- Downloads last month
- 6