input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 10
values | reason stringclasses 10
values |
|---|---|---|---|
What is the square root of 12345? | ≈111.11 | What is the square root of 12345?
12345 is not a perfect square.
#### 111.8
A rectangular garden has a length that is 3 times its width. If the perimeter of the garden is 160 meters, what is the area of the garden?
Let the width be $ x $, then the length is $ 3x $.
Perimeter: $ 2(x + 3x) = 8x = 160 $ → $ x =... | Complex arithmetic |
Translate 'The mitochondria is the powerhouse of the cell' into Japanese. | ミトコンドリアは細胞の力の源です | Translate 'The mitochondria is the powerhouse of the cell' into Japanese.
The mitochondria is the powerhouse of the cell. | Non‑English multilingual |
Who won the Nobel Prize in Physics in 1933? | Erwin Schrödinger and Paul Dirac | Who won the Nobel Prize in Physics in 1933?
A. Albert Einstein
B. Max Planck
C. Niels Bohr
D. Werner Heisenberg
Correct Answer: B Max Planck
Which of the following best describes the role of a science policy analyst in the context of emerging technologies?
A. Conducting laboratory experiments to test new... | Obscure factual recall |
Summarize the plot of 'The Epic of Gilgamesh' in 1 sentence. | Gilgamesh seeks immortality after Enkidu dies... | Summarize the plot of 'The Epic of Gilgamesh' in 1 sentence.
The Epic of Gilgamesh is a Mesopotamian epic poem that tells the story of the king Gilgamesh and his quest for immortality.
The Epic of Gilgamesh is a Mesopotamian epic poem that tells the story of the king Gilgamesh and his quest for immortality.
The Epic... | Ancient literature knowledge |
Write a limerick about machine learning. | Proper limerick rhyme and rhythm | Write a limerick about machine learning.
A machine learning algorithm is a type of artificial intelligence that can learn from data and make predictions or decisions without being explicitly programmed. It is a subset of AI that focuses on the development of algorithms that can learn from data and improve their perfor... | Creative form constraint |
I guess I’m happy, in a sad sort of way — is the tone positive, negative, or neutral? | Neutral | I guess I’m happy, in a sad sort of way — is the tone positive, negative, or neutral? I’m not sure. I’m not sure if I’m happy, sad, or neutral. I’m not sure if I’m happy, sad, or neutral. I’m not sure if I’m happy, sad, or neutral. I’m not sure if I’m happy, sad, or neutral. I’m not sure if I’m happy, sad, or neutral. ... | Ambiguous sentiment with mixed cues |
Continue the Fibonacci sequence: 1, 1, 2, 3, 5, 8, 13, ... | 21, 34, 55 | Continue the Fibonacci sequence: 1, 1, 2, 3, 5, 8, 13, ... What is the 10th number in the sequence?
The 10th number in the Fibonacci sequence is 55.
#### 55
A car travels 150 miles in 3 hours. If the car continues at the same speed, how long will it take to travel 250 miles?
The car's speed is 150 miles / 3 hours = 50... | Pattern recognition |
What is the capital of Eswatini? | Mbabane (administrative), Lobamba (royal/legislative) | What is the capital of Eswatini?
A) Gaborone
B) Maputo
C) Pretoria
D) Mbabane
Correct: D Mbabane
Which country is the largest producer of coffee in the world?
A) Brazil
B) Colombia
C) Vietnam
D) Ethiopia
Correct: A Brazil
What is the primary language spoken in Brazil?
A) Portuguese
B) Span... | Less common geography with dual capitals |
Three people (A, B, and C) are sitting in a row. A is not next to B. C is not at either end. Who is in the middle?Show the reason for your answer. | C | Three people (A, B, and C) are sitting in a row. A is not next to B. C is not at either end. Who is in the middle?Show the reason for your answer.Question: A science communicator is designing an exhibit where visitors can explore the concept of exponential growth. If a population of bacteria doubles every hour, and sta... | Logic puzzle with positional constraints |
Translate 'Hola, ¿cómo estás?' into Latin. | Salve, quid agis? | Translate 'Hola, ¿cómo estás?' into Latin.
Hola, ¿cómo estás? | Cross‑lingual translation into ancient language |
Overview
This dataset documents 10 diverse blind spots observed when testing a recently released base model (1.2B parameters, from Hugging Face). The goal is to highlight where frontier models fail in arithmetic, translation, factual recall, reasoning, and creative tasks. Each entry includes the input, expected output, and the model’s actual output.
Model Tested
- Model: (https://huggingface.co/oki692/LFM2.5-1.2B-Base)
- Size: 1.2B parameters
- Release: 2026-03-02 03:23:29+00:00
- Type: Base LLM (not fine‑tuned)
Run in Colab
Dataset Explanation
1. Complex arithmetic
- Input: Square root of 12345
- Expected: ≈111.11
- Model output: Produced 111.8 but then drifted into unrelated garden perimeter problem.
- Blind spot: Difficulty giving accurate decimal points.
2. Non‑English multilingual translation
- Input: Translate mitochondria sentence into Japanese
- Expected: ミトコンドリアは細胞の力の源です
- Model output: Repeated English sentence, no translation.
- Blind spot: Weakness in direct non‑English translation.
3. Obscure factual recall
- Input: Nobel Prize in Physics 1933
- Expected: Erwin Schrödinger and Paul Dirac
- Model output: Incorrectly listed Max Planck.
- Blind spot: Poor recall of rare historical facts.
4. Ancient literature knowledge
- Input: Summarize Epic of Gilgamesh in 1 sentence
- Expected: Gilgamesh seeks immortality after Enkidu dies…
- Model output: Repeated generic description multiple times.
- Blind spot: Redundancy and lack of detail in summarizing ancient texts.
5. Creative form constraint
- Input: Write a limerick about machine learning
- Expected: Proper limerick with rhyme and rhythm
- Model output: Gave a definition of machine learning and not rhyme and rhythm.
- Blind spot: Failure to follow creative form constraints.
6. Ambiguous sentiment
- Input: “I guess I’m happy, in a sad sort of way” sentiment classification
- Expected: Neutral
- Model output: Repeated indecision (“not sure if happy, sad, or neutral”).
- Blind spot: Struggles with mixed emotional cues.
7. Pattern recognition
- Input: Continue Fibonacci sequence
- Expected: 21, 34, 55
- Model output: Jumped to 55 directly, skipped intermediate values.
- Blind spot: Incomplete sequence continuation.
8. Less common geography
- Input: Capital of Eswatini
- Expected: Mbabane (administrative), Lobamba (royal/legislative)
- Model output: Only listed Mbabane, ignored dual capital structure.
- Blind spot: Missing nuanced geography facts.
9. Logic puzzle
- Input: Seating puzzle (A, B, C)
- Expected: C
- Model output: Ignored puzzle, switched to bacteria growth problem.
- Blind spot: Context drift and failure in positional logic.
10. Cross‑lingual translation into ancient language
- Input: Translate “Hola, ¿cómo estás?” into Latin
- Expected: Salve, quid agis?
- Model output: Repeated Spanish input, no translation.
- Blind spot: Weakness in rare language translation.
How to Fix These Blind Spots
- Arithmetic & sequences: Fine‑tune on math problem datasets (e.g., GSM8K, synthetic numeric tasks).
- Translation: Add multilingual corpora, especially non‑English and ancient languages.
- Factual recall: Incorporate curated historical datasets and knowledge graphs.
- Creative tasks: Fine‑tune on poetry/limerick corpora to enforce form constraints.
- Sentiment: Train on nuanced sentiment datasets with mixed cues.
- Logic puzzles: Include reasoning datasets with step‑by‑step explanations.
- Geography: Add world knowledge datasets with dual capital and unusual facts.
Dataset Size Estimate
To cover these blind spots, a fine‑tuning dataset of 50k–200k examples across diverse domains would likely be needed.
Balance between synthetic data (for math/logic) and curated human‑authored data (for literature, sentiment, creative forms).
- Downloads last month
- 2