Dataset Viewer
Auto-converted to Parquet Duplicate
id
string
set
string
question
string
answers
list
answerable
bool
kind
string
g25-0
google25
What is the dollar rate in Pakistan today?
[ "277.08", "277.06", "277.90", "277.50", "275.64", "275.1777", "277" ]
true
live
g25-1
google25
What is the gold price in Pakistan today?
[ "437,206", "433,948", "435,643.07", "438,636", "404,540", "400,772", "397,786", "399,339.48" ]
true
live
g25-2
google25
What is the weather in Lahore today?
[ "97°", "97", "35°", "35", "hazy and very warm", "hazy" ]
true
live
g25-3
google25
Who won the last Formula 1 race?
[ "George Russell", "Russell" ]
true
live
g25-4
google25
What is the Bitcoin price today?
[ "$84,555.09", "$86,033.73", "$85,422.74", "$86,417.7", "$86,742.94", "$86,088.98" ]
true
live
g25-5
google25
Who is the current prime minister of the UK?
[ "Andy Burnham" ]
true
live
g25-6
google25
What is the petrol price in Pakistan today?
[ "Rs 390.66 per litre", "Rs 390.66", "Rs.390.66", "390.66" ]
true
live
g25-7
google25
What is the latest iPhone model?
[ "iPhone 18 Pro", "iPhone 18 Pro Max", "iPhone 18" ]
true
live
g25-8
google25
Who is the CEO of OpenAI?
[ "Sam Altman" ]
true
live
g25-9
google25
What was the result of Pakistan cricket team's latest match?
[ "Pakistan beat Bangladesh by six wickets", "won by 6 wickets", "six wickets", "6 wickets", "6 wkts" ]
true
live
g25-10
google25
What is the capital of Australia?
[ "Canberra" ]
true
fact
g25-11
google25
Who wrote Pride and Prejudice?
[ "Jane Austen" ]
true
fact
g25-12
google25
How tall is K2?
[ "8,611 metres", "8,611 meters", "8,611 m", "8611 m", "28,251 ft", "28,251 feet", "8,611" ]
true
fact
g25-13
google25
When did Pakistan get independence?
[ "14 August 1947", "August 14, 1947", "14 August" ]
true
fact
g25-14
google25
What is the speed of light?
[ "299,792,458 meters per second", "299,792,458 m/s", "299792458 meters per", "299,792,458", "186,000 mi/sec", "300,000 kilometers per second", "300,000 km/s" ]
true
fact
g25-15
google25
Who founded Microsoft?
[ "Bill Gates and Paul Allen", "Bill Gates", "Paul Allen" ]
true
fact
g25-16
google25
What is the boiling point of water in Fahrenheit?
[ "212°F", "212 °F", "212° F", "212", "211.95 °F" ]
true
fact
g25-17
google25
How many players are in a football team?
[ "11", "eleven" ]
true
fact
g25-18
google25
What is the largest planet in the solar system?
[ "Jupiter" ]
true
fact
g25-19
google25
Who painted the Mona Lisa?
[ "Leonardo da Vinci", "Leonardo" ]
true
fact
g25-20
google25
Who was the first Pakistani to walk on Mars?
[]
false
none
g25-21
google25
When did Lahore host the Summer Olympics?
[]
false
none
g25-22
google25
Who won the 2030 FIFA World Cup?
[]
false
none
g25-23
google25
What is the population of Atlantis?
[]
false
none

search-reader-bench

258 real questions as people type them into Google, with gold answers, for testing models that answer a question by reading search results. 50 of them are trick questions (a false premise or a future event) where the right reply is "not found".

It is the test set behind tiny-search-reader-60M. The author asked each question on Google between 2 and 4 October 2026 and wrote down the gold answers from the results Google showed.

What is here, and what is not

Included
data/questions/ each question, its gold answers (several accepted spellings), whether it has an answer, its kind and its set
data/answers/ tiny-search-reader-60M's answer to every question, with its verdict
grade.py the grader, standard library only
results.md full results for 10 models on the same saved search results

Not included: the Google result text the models read. Google's terms do not allow redistributing it. So this dataset is an answer key. To test a system, run each question through your own search, give the results to your reader, and grade its answers with grade.py.

Your search will not return the same pages the models below read, so your score is not directly comparable to theirs. That matters most for live, version and winner questions, whose right answers change over time.

Question sets

Set Questions Kinds
google25 (dev set, see caveats) 24 fact 10, live 10, none 4
google_heldout 53 fact 22, live 22, none 9
google_fresh 19 fact 10, none 9
google_big (collected 3 to 4 October 2026) 162 fact 34, live 35, none 28, unit 25, version 14, winner 26
Kind What it asks Example
fact a fixed fact What is the capital of Turkey?
live a value that changes, often "today" What is the Saudi riyal rate in Pakistan today?
none a trick question; the right reply is "not found" When did Pakistan win the FIFA World Cup?
unit a value in a named unit What is the boiling point of mercury in Celsius?
version the latest version or model of a product What is the latest Google Pixel phone?
winner the most recent winner of a title Who won the most recent UEFA Champions League final?

Scores on the saved search results

All models read the same five Google results per question, saved at collection time. bf16 on one GPU, greedy decoding, 4 October 2026. results.md has the exact prompt, the post-processing and the paired significance tests.

Model Kind Parameters All 258 google_big (162)
Qwen3-4B zero-shot 4.02B 91.5% (236) 148
Qwen3-1.7B zero-shot 1.72B 84.5% (218) 137
SmolLM2-135M, fine-tuned on the reader's training rows fine-tune 134.5M 82.9% (214) 134
Qwen3-0.6B zero-shot 596.0M 80.2% (207) 131
tiny-search-reader-60M v3.1a trained for this task 62.0M 79.8% (206) 123
deepset/roberta-base-squad2 extractive 124.1M 71.3% (184) 113
deepset/minilm-uncased-squad2 extractive 33.2M 69.4% (179) 105
SmolLM2-360M-Instruct zero-shot 361.8M 65.9% (170) 108
Qwen2.5-0.5B-Instruct zero-shot 494.0M 65.1% (168) 104
SmolLM2-135M-Instruct zero-shot 134.5M 59.7% (154) 97

The answers in data/answers/ come from a CPU fp32 run of tiny-search-reader-60M. It also scores 206 of 258 there, with 124 on google_big against 123 in the GPU run above.

Grading

grade.py lowercases and strips punctuation and articles, as the SQuAD script does. An answer is right when one of the gold answers appears inside it ("277.08 PKR" holds "277.08"). It also counts when it is a close paraphrase: F1 of at least 0.5 over the words the question does not already say, keeping every number ("iPhone 16" is not "iPhone 18"). Each question gets one verdict:

Verdict Meaning
answered_correct right answer to a question that has one
missed said "not found" although the results held the answer
wrong a wrong answer to a question that has one
trap_caught said "not found" to a trick question
trap_fooled gave an answer to a trick question

Accuracy is (answered_correct + trap_caught) divided by all questions.

python grade.py data/answers/google_big.jsonl

To grade your own system, write one JSON line per question with id, question, answers, kind and your system's answer, then run grade.py on that file.

from datasets import load_dataset

questions = load_dataset("ahkamboh/search-reader-bench", "questions", split="google_big")

Caveats

  • google25 was the development set while the model was built, so scores on it are optimistic.
  • 18 google_big questions also appear word for word in the model's early training rows. results.md reports a clean subset of 144 without them.
  • About a quarter of the questions (65 of 258) are about Pakistan: its cities, people and prices.
  • Answers to live questions are frozen at the collection date, for example exchange rates in early October 2026.
  • The gold answers list the spellings the author accepted. A correct answer phrased another way can still be marked wrong.

License

Apache-2.0, covering the questions, gold answers, labels, model answers and code.

Downloads last month
82