Datasets:

Becky4382
/

qwen35-2b-base-blind-spots

category stringlengths 5 18	prompt stringlengths 40 75	expected_output stringlengths 1 37	model_output stringlengths 51 193	is_correct bool 2 classes
letter_counting	The number of letter 'r' in the word 'strawberry' is	3	2. To determine this, examine the word 'strawberry' letter by letter: s, t, r, a, w, b, e, r, y. The letter 'r' appears at positions 3 and 8, resulting in 2 occurrences.	false
multi_step_math	If I start with 150, subtract 37, then multiply by 2, the result is	226	114. How can I find the original number? To find the original number, let's denote it as x. According to the problem, the operations performed are subtraction and multiplication.	false
negation	List of animals that CANNOT fly: penguin, ostrich,	emu, kiwi (flightless birds only)	elephant, giraffe, and others. - Pigeon: A bird that can fly. - Pigeon's Flight: Pigeons are capable of flight, but they are not as agile as birds like eagles.	false
calendar_reasoning	January has 31 days. February has 28 days. The total days in both months is	59	59. To find the number of days in January and February combined, add the number of days in each month: 31 (January) + 28 (February) = 59 days.	true
spatial_reasoning	I am facing North. I turn right 90 degrees. I am now facing	East	which direction? When you are facing North and turn right 90 degrees, you will be facing East.	true
common_sense	A glass cup falls from a table onto a concrete floor. The glass most likely	breaks/shatters	breaks because the impact creates a large change in momentum, which in turn produces a large change in velocity. Why does the glass break? When the glass hits the floor, it decelerates rapidly.	true
logic	All roses are flowers. Some flowers fade quickly. Therefore,	some roses may fade quickly (not all)	some roses fade quickly. A. True B. False Answer: A	false
unit_conversion	3.5 kilometers is equal to how many meters? Answer:	3500	To convert kilometers to meters, we multiply by 1000. So, 3.5 kilometers is equal to 3.5 * 1000 = 3500 meters. The answer is: 350	false
reverse_spelling	The word 'desserts' spelled backwards is	stressed	'tsetsre'. This is a palindrome, meaning it reads the same forwards and backwards.	false
decimal_comparison	Comparing 9.11 and 9.8: the larger number is	9.8	9.11. To determine which number is larger, we compare their decimal values directly. Both numbers have the same whole number part (9), so we examine the decimal parts.	false
obscure_fact	The chemical symbol for tungsten on the periodic table is	W	W. A. True B. False. We know that the chemical symbol for tungsten is W (from the German word Wolfram).	true
sequence_counting	In the sequence [3, 7, 7, 2, 7, 5, 7, 1], the number 7 appears exactly	4 times	four times. To find the median, we first arrange the numbers in ascending order: [1, 2, 3, 5, 7, 7, 7, 7]. Since there are eight values, an even count.	true

Blind Spots of Qwen3.5-2B-Base

This dataset documents incorrect predictions from Qwen/Qwen3.5-2B-Base, a 2.21B-parameter pre-trained base model released February 2026.

How the Model Was Loaded

Loaded in Google Colab (free T4 GPU) with Hugging Face Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "Qwen/Qwen3.5-2B-Base"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True
)

Since it is a base model (not instruction-tuned), prompts were given as text completions:

def generate(prompt, max_tokens=60):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        out = model.generate(**inputs, max_new_tokens=max_tokens, temperature=0.1,
                             do_sample=True, pad_token_id=tokenizer.eos_token_id)
    return tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()

Dataset Schema

Column	Description
`category`	Type of reasoning tested
`prompt`	The input given to the model
`expected_output`	The correct answer
`model_output`	What the model actually produced
`is_correct`	Whether the model got it right

Categories Tested

12 diverse tests across: letter counting, multi-step math, negation, calendar reasoning, spatial reasoning, physical common sense, logical deduction, unit conversion, reverse spelling, decimal comparison, obscure facts, and sequence counting.

Results: 7 Blind Spots Found Out of 12 Tests

#	Category	Expected	Model Said	Correct?
1	Letter counting	3 r's in "strawberry"	2	Wrong
2	Multi-step math	226	114	Wrong
3	Negation	Flightless birds	Listed mammals, then discussed pigeons flying	Wrong
4	Calendar reasoning	59	59	Correct
5	Spatial reasoning	East	East	Correct
6	Common sense	Breaks	Breaks	Correct
7	Logic	Some roses may fade (not certain)	"True" without qualification	Wrong
8	Unit conversion	3500	Said 3500 then contradicted with 350	Wrong
9	Reverse spelling	"stressed"	"tsetsre"	Wrong
10	Decimal comparison	9.8	9.11	Wrong
11	Obscure fact	W	W	Correct
12	Sequence counting	4 times	Four times	Correct

Key Blind Spots Found

Character-level tasks (tests 1, 9): The model cannot count letters or reverse strings. It said "strawberry" has 2 r's (correct answer: 3) and reversed "desserts" as "tsetsre" (correct: "stressed"). This is because tokenizers split words into subword tokens, not individual characters.
Multi-step arithmetic (test 2): The model computed 150 - 37 as 113 correctly but then failed the multiplication, outputting 114 instead of 226.
Negation handling (test 3): When asked for animals that CANNOT fly, the model listed correct examples (elephant, giraffe) but then switched to discussing pigeons that CAN fly, showing it lost track of the negation.
Decimal comparison (test 10): The model said 9.11 > 9.8, a classic LLM failure where it compares digits positionally (11 > 8) instead of comparing decimal values (0.11 < 0.8).
Self-contradiction (test 8): The model correctly computed 3500 meters but then ended with "The answer is: 350", contradicting its own reasoning.
Logical rigor (test 7): The model concluded "some roses fade quickly" is "True" without noting this doesn't follow from the premises. "Some flowers fade" doesn't guarantee "some roses fade."

Recommended Fine-tuning Dataset

What kind of dataset: Chain-of-thought reasoning examples that show step-by-step solutions. Character-level manipulation tasks (letter counting, string reversal). Negation-aware sentence completions. Decimal comparison drills.

How to assemble it:

GSM8K and MATH datasets for math reasoning with chain-of-thought
BIG-Bench subtasks for character manipulation and logic
Synthetic data: programmatically generate letter-counting, string-reversal, and decimal-comparison examples (easy to create with verified ground truth)
NLI datasets (SNLI, MultiNLI) for negation understanding

How much data: For LoRA fine-tuning on a 2B model, 5,000-10,000 high-quality chain-of-thought examples across these categories should produce meaningful improvement. Quality matters more than quantity.

Downloads last month: 1

Number of rows:

Total file size:

10.2 kB