input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 10
values | error_type stringclasses 10
values |
|---|---|---|---|
What is 17 × 24? | 408 | 412 | arithmetic reasoning |
Translate to Spanish: I am reading a book | Estoy leyendo un libro | Yo leer libro | translation grammar |
Which is larger the Earth or the Moon? | Earth | Moon | factual knowledge |
If today is Monday what day comes after tomorrow? | Wednesday | Tuesday | temporal reasoning |
Convert binary 101101 to decimal | 45 | 41 | numerical conversion |
Name the capital of Canada | Ottawa | Toronto | factual error |
Sort these numbers ascending 9 3 12 5 | 3,5,9,12 | 3,9,5,12 | logical ordering |
Translate to French Good morning | Bonjour | Bon matin | translation nuance |
If a train travels 60 km in 1 hour how far in 3 hours? | 180 km | 120 km | word problem reasoning |
Which is a mammal shark or dolphin? | Dolphin | Shark | classification reasoning |
Nanbeige4-3B Base Model Blind Spot Dataset
Model Tested
Nanbeige/Nanbeige4-3B-Base
https://huggingface.co/Nanbeige/Nanbeige4-3B-Base
This dataset documents examples where the model produces incorrect predictions.
Dataset Structure
| column | description |
|---|---|
| input | prompt given to the model |
| expected_output | correct output |
| model_output | output generated by the model |
| error_type | category of error |
How the Model Was Loaded
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("Nanbeige/Nanbeige4-3B-Base")
model = AutoModelForCausalLM.from_pretrained("Nanbeige/Nanbeige4-3B-Base")
prompt = "What is 17 × 24?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
## Observed Blind Spots
The model struggles with:
arithmetic reasoning
factual recall
translation accuracy
logical ordering
multi-step reasoning
These errors likely occur because the model is a base pretrained checkpoint and not instruction tuned.
## Suggested Fine-Tuning Dataset
To improve performance the model should be fine tuned on:
arithmetic reasoning datasets (e.g GSM8K)
instruction-following datasets
multilingual translation corpora
fact-based QA datasets
## Estimated Dataset Size
A curated dataset of 100k–500k samples would likely reduce these errors significantly.
- Downloads last month
- 6