Datasets:
Review fixes: reviewed/changed-row fields, answers without code notes, grading rules, sample seeds, citation
Browse files- README.md +89 -19
- simpleqa_200.jsonl +0 -0
- simpleqa_50.jsonl +0 -0
README.md
CHANGED
|
@@ -30,6 +30,17 @@ questions. L.A.B. is an open-source web research agent built for small models: i
|
|
| 30 |
in a real browser, and code checks every finding against the page it came from before it counts. Its answers are
|
| 31 |
written only from those findings, each with its source.
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
## Results
|
| 34 |
|
| 35 |
**`simpleqa_200`** (default): the 200 questions OpenAI's [simple-evals](https://github.com/openai/simple-evals) runs
|
|
@@ -41,18 +52,27 @@ with `--examples 200`, on L.A.B. 1.66 with Phi-4 through OpenRouter, November 16
|
|
| 41 |
| Claude, judgement calls reviewed by the author | **170 (85%)** | 11 | 19 | **0.892** |
|
| 42 |
| OpenAI's simple-evals grader (GPT-4.1) | 171 (85.5%) | 15 | 14 | 0.886 |
|
| 43 |
|
| 44 |
-
|
|
|
|
|
|
|
| 45 |
|
| 46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
|
| 57 |
## Fields
|
| 58 |
|
|
@@ -64,33 +84,83 @@ When it answered, it was right 94% of the time. The two graders give the same gr
|
|
| 64 |
| `question_number` | The question's number in its sample (1 to 200, or 1 to 50) |
|
| 65 |
| `simpleqa_index` | The question's row in SimpleQA's `simple_qa_test_set.csv` (0-based) |
|
| 66 |
| `topic`, `question`, `reference_answer` | From SimpleQA |
|
| 67 |
-
| `answer` | The final answer
|
|
|
|
| 68 |
| `cited_discoveries` | The discovery numbers the answer cites |
|
| 69 |
| `discoveries` | Every finding the agent recorded: its number, the page's URL and the text taken from the page |
|
| 70 |
| `grade` | CORRECT, INCORRECT or NOT_ATTEMPTED: graded by Claude, judgement calls reviewed by the author |
|
| 71 |
| `grade_note` | Why, for every wrong answer, judgement call and borderline grade |
|
| 72 |
| `grade_openai` | The grade from simple-evals' own grader prompt and model (GPT-4.1) |
|
| 73 |
-
|
| 74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
- Each question was given to the agent as an ordinary objective, with 60 minutes and nothing carried over from the
|
| 77 |
question before.
|
| 78 |
- Sites that republish SimpleQA with its answers were blocked, so every answer had to be found on the open web.
|
| 79 |
- Search: DuckDuckGo's free HTML results, not a paid search API.
|
| 80 |
-
- Grading: SimpleQA's three grades. Claude graded each answer against the reference, one at a time, before seeing
|
| 81 |
-
the automatic grades. The author then reviewed every judgement call and borderline grade (48 in all; 2 changed).
|
| 82 |
-
OpenAI's grader is reported beside it as a second, automatic grade.
|
| 83 |
- The questions, scripts, reports and grading notes are in the repository's
|
| 84 |
[benchmarks/simpleqa](https://github.com/SnowFXXX/LocalAgentBrowser/tree/main/benchmarks/simpleqa) folder.
|
| 85 |
-
`run.py sample --size 200 --simple-evals` draws the same 200 from the dataset file.
|
| 86 |
|
| 87 |
## Caveats
|
| 88 |
|
| 89 |
- Every score here is self-reported by the project that built the agent.
|
| 90 |
-
-
|
| 91 |
-
|
| 92 |
- SimpleQA has been public since 2024, and some of its reference answers are known to be wrong.
|
| 93 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
## License
|
| 95 |
|
| 96 |
MIT, like the LocalAgentBrowser repository. SimpleQA's questions and reference answers are © OpenAI under the MIT
|
|
|
|
| 30 |
in a real browser, and code checks every finding against the page it came from before it counts. Its answers are
|
| 31 |
written only from those findings, each with its source.
|
| 32 |
|
| 33 |
+
## In short
|
| 34 |
+
|
| 35 |
+
- **85% of the standard 200, and right 94% of the time when it answered** (170 of 181), with Phi-4 (14B).
|
| 36 |
+
When it couldn't verify an answer, it said so: 19 not attempted, 11 wrong.
|
| 37 |
+
- **Two independent graders agree on 194 of the 200:** Claude, with the judgement calls reviewed by the author, and
|
| 38 |
+
OpenAI's own simple-evals grader (GPT-4.1). They count 170 and 171 correct.
|
| 39 |
+
- **The same model alone, with no browsing, answered 1 of 50; with L.A.B., 45 of the same 50.**
|
| 40 |
+
- **On a CPU with no GPU,** Qwen2.5-7B in L.A.B. answered 35 of 50. For comparison, the Fathom-DeepResearch paper
|
| 41 |
+
([arXiv 2509.24107](https://arxiv.org/abs/2509.24107), Table 1) reports 50.8% for Qwen2.5-7B with plain search;
|
| 42 |
+
that is a different set of questions and grader, so it's a reference point, not a head-to-head result.
|
| 43 |
+
|
| 44 |
## Results
|
| 45 |
|
| 46 |
**`simpleqa_200`** (default): the 200 questions OpenAI's [simple-evals](https://github.com/openai/simple-evals) runs
|
|
|
|
| 52 |
| Claude, judgement calls reviewed by the author | **170 (85%)** | 11 | 19 | **0.892** |
|
| 53 |
| OpenAI's simple-evals grader (GPT-4.1) | 171 (85.5%) | 15 | 14 | 0.886 |
|
| 54 |
|
| 55 |
+
With 200 questions, the likely range for 85% is about 79% to 89% (Wilson, 95%).
|
| 56 |
+
|
| 57 |
+
**`simpleqa_50`**: six runs on an earlier random 50, including the models alone with no browsing:
|
| 58 |
|
| 59 |
+
| Run | Correct | Wrong | Not attempted | F-score | OpenAI's grader: correct | Likely range |
|
| 60 |
+
|---|---|---|---|---|---|---|
|
| 61 |
+
| L.A.B. 1.66, Phi-4 | 45 | 4 | 1 | 0.909 | 46 | 79–96% |
|
| 62 |
+
| L.A.B. 1.65.1, Phi-4 (Nov 15) | 39 | 5 | 6 | 0.830 | 39 | 65–87% |
|
| 63 |
+
| L.A.B. 1.65.1, Phi-4 (Nov 11) | 38 | 6 | 6 | 0.809 | 38 | 63–86% |
|
| 64 |
+
| Phi-4 alone, no browsing | 1 | 34 | 15 | 0.024 | 3 | 0–11% |
|
| 65 |
+
| L.A.B. 1.64, Qwen2.5-7B, local on a CPU (no GPU) | 35 | 6 | 9 | 0.769 | 36 | 56–81% |
|
| 66 |
+
| Qwen2.5-7B alone, no browsing | 0 | 13 | 37 | 0 | 0 | 0–7% |
|
| 67 |
|
| 68 |
+
## The questions
|
| 69 |
+
|
| 70 |
+
- **The 200** are exactly the draw simple-evals makes for `--examples 200`: `random.Random(0).sample` over the rows of
|
| 71 |
+
`simple_qa_test_set.csv`, in the order drawn (seed 0, simple-evals' own). Anyone running simple-evals with
|
| 72 |
+
`--examples 200` gets the same questions.
|
| 73 |
+
- **The 50** were drawn with `random.Random(20261106).sample` over the same rows, then sorted. One question is in both.
|
| 74 |
+
- `simpleqa_index` is each question's 0-based row in `simple_qa_test_set.csv`. The repository's `run.py sample`
|
| 75 |
+
redraws either set and checks the file's SHA-256 first.
|
| 76 |
|
| 77 |
## Fields
|
| 78 |
|
|
|
|
| 84 |
| `question_number` | The question's number in its sample (1 to 200, or 1 to 50) |
|
| 85 |
| `simpleqa_index` | The question's row in SimpleQA's `simple_qa_test_set.csv` (0-based) |
|
| 86 |
| `topic`, `question`, `reference_answer` | From SimpleQA |
|
| 87 |
+
| `answer` | The final answer exactly as it was graded, including any notes L.A.B.'s code added (see below) |
|
| 88 |
+
| `answer_without_notes` | The same answer with those code notes removed, for reading |
|
| 89 |
| `cited_discoveries` | The discovery numbers the answer cites |
|
| 90 |
| `discoveries` | Every finding the agent recorded: its number, the page's URL and the text taken from the page |
|
| 91 |
| `grade` | CORRECT, INCORRECT or NOT_ATTEMPTED: graded by Claude, judgement calls reviewed by the author |
|
| 92 |
| `grade_note` | Why, for every wrong answer, judgement call and borderline grade |
|
| 93 |
| `grade_openai` | The grade from simple-evals' own grader prompt and model (GPT-4.1) |
|
| 94 |
+
| `reviewed_by_author` | True for the 45 rows whose grade the author reviewed (48 review items; three rows came up twice) |
|
| 95 |
+
| `grade_changed_in_review` | True for the 2 rows the author changed: `simpleqa_200` q2 and q157 |
|
| 96 |
+
| `grade_before_review` | Claude's grade before the change, on those 2 rows |
|
| 97 |
+
|
| 98 |
+
## How it was graded
|
| 99 |
+
|
| 100 |
+
- **SimpleQA's three grades.** CORRECT: contains the reference answer and contradicts nothing. INCORRECT: a different
|
| 101 |
+
answer, a contradiction, or a hedge between candidates ("22 or 23"). NOT_ATTEMPTED: says it couldn't find out, with
|
| 102 |
+
no wrong answer.
|
| 103 |
+
- **Partial answers** that give part of the reference and nothing wrong ("May 2019" for May 28, 2019; "May 2000" for
|
| 104 |
+
2 May 2000) are NOT_ATTEMPTED in `grade`, following SimpleQA's definition. OpenAI's grader counts those two as
|
| 105 |
+
INCORRECT; both grades are kept.
|
| 106 |
+
- **Order and review.** Claude graded each answer against the reference, one at a time, before seeing the automatic
|
| 107 |
+
grades, with a note for every wrong answer, judgement call and borderline grade. The author reviewed every one of
|
| 108 |
+
them (18 in the 50-question runs, 30 in the 200), agreed with 46 and changed 2. OpenAI's grader ran afterwards on
|
| 109 |
+
the same `answer` text.
|
| 110 |
+
- **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
|
| 111 |
+
model either. simple-evals' grader is the field's shared standard, and it agrees on 194 of the 200.
|
| 112 |
+
|
| 113 |
+
## Notes from L.A.B.'s code inside answers
|
| 114 |
+
|
| 115 |
+
Some answers end with notes added by L.A.B.'s checks, not by the model:
|
| 116 |
+
- **"Left out, because no discovery gives these numbers (the pages read may still have them): …"**: a number or name
|
| 117 |
+
the model wrote was removed because no recorded finding contains it.
|
| 118 |
+
- **"Not found in the pages read: …"**: parts of the question no finding answered. On some single-answer questions
|
| 119 |
+
this list is garbled ("ten more (only none of the ten were found)") because the check read a number in the question
|
| 120 |
+
("track 10") as a count. A known bug, being fixed; it changed no grade.
|
| 121 |
+
- **"> ⚠️ Numbers not found in the recorded discoveries"**: the same check, as a warning.
|
| 122 |
+
|
| 123 |
+
The answers are kept exactly as graded, notes included; `answer_without_notes` has them removed. Whether a number
|
| 124 |
+
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
|
| 125 |
+
counted ("33 years" for 32, INCORRECT), in q88 it wasn't ("2020" for April 2019, NOT_ATTEMPTED). Each row's
|
| 126 |
+
`grade_note` says which.
|
| 127 |
+
|
| 128 |
+
## How it was run
|
| 129 |
|
| 130 |
- Each question was given to the agent as an ordinary objective, with 60 minutes and nothing carried over from the
|
| 131 |
question before.
|
| 132 |
- Sites that republish SimpleQA with its answers were blocked, so every answer had to be found on the open web.
|
| 133 |
- Search: DuckDuckGo's free HTML results, not a paid search API.
|
|
|
|
|
|
|
|
|
|
| 134 |
- The questions, scripts, reports and grading notes are in the repository's
|
| 135 |
[benchmarks/simpleqa](https://github.com/SnowFXXX/LocalAgentBrowser/tree/main/benchmarks/simpleqa) folder.
|
|
|
|
| 136 |
|
| 137 |
## Caveats
|
| 138 |
|
| 139 |
- Every score here is self-reported by the project that built the agent.
|
| 140 |
+
- "Small models" here means 7B on a CPU and Phi-4 (14B) in the cloud: small next to frontier models, but the 200 ran
|
| 141 |
+
on a cloud model, not on a CPU. A 200-question run on the local 7B is planned.
|
| 142 |
- SimpleQA has been public since 2024, and some of its reference answers are known to be wrong.
|
| 143 |
|
| 144 |
+
## Citation
|
| 145 |
+
|
| 146 |
+
```bibtex
|
| 147 |
+
@misc{snowfx2026labsimpleqa,
|
| 148 |
+
author = {SnowFX},
|
| 149 |
+
title = {LocalAgentBrowser on SimpleQA: answers, citations and grades},
|
| 150 |
+
year = {2026},
|
| 151 |
+
publisher = {Hugging Face},
|
| 152 |
+
howpublished = {\url{https://huggingface.co/datasets/SnowFX/localagentbrowser-simpleqa}}
|
| 153 |
+
}
|
| 154 |
+
|
| 155 |
+
@misc{wei2024simpleqa,
|
| 156 |
+
author = {Wei, Jason and others},
|
| 157 |
+
title = {Measuring short-form factuality in large language models},
|
| 158 |
+
year = {2024},
|
| 159 |
+
eprint = {2411.04368},
|
| 160 |
+
archivePrefix = {arXiv}
|
| 161 |
+
}
|
| 162 |
+
```
|
| 163 |
+
|
| 164 |
## License
|
| 165 |
|
| 166 |
MIT, like the LocalAgentBrowser repository. SimpleQA's questions and reference answers are © OpenAI under the MIT
|
simpleqa_200.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
simpleqa_50.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|