SnowFX's picture
Card wording: give the not-attempted and wrong counts; "two graders", not "independent"
c7c224e verified
|
Raw History Blame Contribute Delete
8.9 kB
metadata
license: mit
language:
  - en
pretty_name: LocalAgentBrowser on SimpleQA
size_categories:
  - n<1K
task_categories:
  - question-answering
tags:
  - simpleqa
  - web-agent
  - evaluation
  - grounding
  - citations
  - small-models
configs:
  - config_name: simpleqa_200
    data_files: simpleqa_200.jsonl
    default: true
  - config_name: simpleqa_50
    data_files: simpleqa_50.jsonl

LocalAgentBrowser on SimpleQA

Every answer, citation and grade from LocalAgentBrowser (L.A.B.) runs on SimpleQA, OpenAI's benchmark of short, hard factual questions. L.A.B. is an open-source web research agent built for small models: it searches DuckDuckGo, reads pages in a real browser, and code checks every finding against the page it came from before it counts. Its answers are written only from those findings, each with its source.

In short

  • 85% of the standard 200, and right 94% of the time when it answered (170 of 181), with Phi-4 (14B). 19 times it said it couldn't verify an answer rather than guess; 11 answers were wrong.
  • Two graders agree on 194 of the 200: Claude, with the judgement calls reviewed by the author, and OpenAI's own simple-evals grader (GPT-4.1). They count 170 and 171 correct.
  • The same model alone, with no browsing, answered 1 of 50; with L.A.B., 45 of the same 50.
  • On a CPU with no GPU, Qwen2.5-7B in L.A.B. answered 35 of 50. For comparison, the Fathom-DeepResearch paper (arXiv 2509.24107, Table 1) reports 50.8% for Qwen2.5-7B with plain search; that is a different set of questions and grader, so it's a reference point, not a head-to-head result.

Results

simpleqa_200 (default): the 200 questions OpenAI's simple-evals runs with --examples 200, on L.A.B. 1.66 with Phi-4 through OpenRouter, November 16, 2026. The run took 142 minutes and 2,762 model calls, for $0.58.

Grader Correct Wrong Not attempted F-score
Claude, judgement calls reviewed by the author 170 (85%) 11 19 0.892
OpenAI's simple-evals grader (GPT-4.1) 171 (85.5%) 15 14 0.886

With 200 questions, the likely range for 85% is about 79% to 89% (Wilson, 95%).

simpleqa_50: six runs on an earlier random 50, including the models alone with no browsing:

Run Correct Wrong Not attempted F-score OpenAI's grader: correct Likely range
L.A.B. 1.66, Phi-4 45 4 1 0.909 46 79–96%
L.A.B. 1.65.1, Phi-4 (Nov 15) 39 5 6 0.830 39 65–87%
L.A.B. 1.65.1, Phi-4 (Nov 11) 38 6 6 0.809 38 63–86%
Phi-4 alone, no browsing 1 34 15 0.024 3 0–11%
L.A.B. 1.64, Qwen2.5-7B, local on a CPU (no GPU) 35 6 9 0.769 36 56–81%
Qwen2.5-7B alone, no browsing 0 13 37 0 0 0–7%

The questions

  • The 200 are exactly the draw simple-evals makes for --examples 200: random.Random(0).sample over the rows of simple_qa_test_set.csv, in the order drawn (seed 0, simple-evals' own). Anyone running simple-evals with --examples 200 gets the same questions.
  • The 50 were drawn with random.Random(20261106).sample over the same rows, then sorted. One question is in both.
  • simpleqa_index is each question's 0-based row in simple_qa_test_set.csv. The repository's run.py sample redraws either set and checks the file's SHA-256 first.

Fields

Field What it is
run Which run the row belongs to
setting L.A.B. with a cloud or local model, or the model alone
model, lab_version The model, and the L.A.B. version (none for the model alone)
question_number The question's number in its sample (1 to 200, or 1 to 50)
simpleqa_index The question's row in SimpleQA's simple_qa_test_set.csv (0-based)
topic, question, reference_answer From SimpleQA
answer The final answer exactly as it was graded, including any notes L.A.B.'s code added (see below)
answer_without_notes The same answer with those code notes removed, for reading
cited_discoveries The discovery numbers the answer cites
discoveries Every finding the agent recorded: its number, the page's URL and the text taken from the page
grade CORRECT, INCORRECT or NOT_ATTEMPTED: graded by Claude, judgement calls reviewed by the author
grade_note Why, for every wrong answer, judgement call and borderline grade
grade_openai The grade from simple-evals' own grader prompt and model (GPT-4.1)
reviewed_by_author True for the 45 rows whose grade the author reviewed (48 review items; three rows came up twice)
grade_changed_in_review True for the 2 rows the author changed: simpleqa_200 q2 and q157
grade_before_review Claude's grade before the change, on those 2 rows

How it was graded

  • SimpleQA's three grades. CORRECT: contains the reference answer and contradicts nothing. INCORRECT: a different answer, a contradiction, or a hedge between candidates ("22 or 23"). NOT_ATTEMPTED: says it couldn't find out, with no wrong answer.
  • Partial answers that give part of the reference and nothing wrong ("May 2019" for May 28, 2019; "May 2000" for 2 May 2000) are NOT_ATTEMPTED in grade, following SimpleQA's definition. OpenAI's grader counts those two as INCORRECT; both grades are kept.
  • Order and review. Claude graded each answer against the reference, one at a time, before seeing the automatic grades, with a note for every wrong answer, judgement call and borderline grade. The author reviewed every one of them (18 in the 50-question runs, 30 in the 200), agreed with 46 and changed 2. OpenAI's grader ran afterwards on the same answer text.
  • Why a second grader: L.A.B.'s principle is not taking a model's word for it, so the grades shouldn't rest on one model either. simple-evals' grader is the field's shared standard, and it agrees on 194 of the 200.

Notes from L.A.B.'s code inside answers

Some answers end with notes added by L.A.B.'s checks, not by the model:

  • "Left out, because no discovery gives these numbers (the pages read may still have them): …": a number or name the model wrote was removed because no recorded finding contains it.
  • "Not found in the pages read: …": parts of the question no finding answered. On some single-answer questions this list is garbled ("ten more (only none of the ten were found)") because the check read a number in the question ("track 10") as a count. A known bug, being fixed; it changed no grade.
  • "> ⚠️ Numbers not found in the recorded discoveries": the same check, as a warning.

The answers are kept exactly as graded, notes included; answer_without_notes has them removed. Whether a number inside a "Left out" note counts as the answer is the author's judgement call row by row: in q2 of the 200 it was counted ("33 years" for 32, INCORRECT), in q88 it wasn't ("2020" for April 2019, NOT_ATTEMPTED). Each row's grade_note says which.

How it was run

  • Each question was given to the agent as an ordinary objective, with 60 minutes and nothing carried over from the question before.
  • Sites that republish SimpleQA with its answers were blocked, so every answer had to be found on the open web.
  • Search: DuckDuckGo's free HTML results, not a paid search API.
  • The questions, scripts, reports and grading notes are in the repository's benchmarks/simpleqa folder.

Caveats

  • Every score here is self-reported by the project that built the agent.
  • "Small models" here means 7B on a CPU and Phi-4 (14B) in the cloud: small next to frontier models, but the 200 ran on a cloud model, not on a CPU. A 200-question run on the local 7B is planned.
  • SimpleQA has been public since 2024, and some of its reference answers are known to be wrong.

Citation

@misc{snowfx2026labsimpleqa,
  author       = {SnowFX},
  title        = {LocalAgentBrowser on SimpleQA: answers, citations and grades},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SnowFX/localagentbrowser-simpleqa}}
}

@misc{wei2024simpleqa,
  author        = {Wei, Jason and others},
  title         = {Measuring short-form factuality in large language models},
  year          = {2024},
  eprint        = {2411.04368},
  archivePrefix = {arXiv}
}

License

MIT, like the LocalAgentBrowser repository. SimpleQA's questions and reference answers are © OpenAI under the MIT license. The discoveries texts are short excerpts from the pages listed beside them, and remain under those pages' own terms.