Publish verified equivalent article references
#2
by gehuageyan - opened
- README.md +5 -3
- data/factual-v1.jsonl +0 -0
- definitions.tar.gz +2 -2
- definitions/datasets/factual-v1.json +0 -0
- definitions/reference-equivalence.json +38 -0
- definitions/registry.json +1 -1
- manifest.json +21 -13
README.md
CHANGED
|
@@ -46,7 +46,7 @@ The `factual-v1` configuration contains every question in the pinned 1,000-row p
|
|
| 46 |
|
| 47 |
Publisher row order, question text and reference answer text are retained. Case IDs use `sqv-` followed by the publisher's `original_index`, preserving the link to OpenAI SimpleQA. Supporting URLs are split from the CSV field, deduplicated in order, and stripped of unmatched trailing closing parentheses. Topic and answer-type labels are retained. The publisher's `multi_step` and `requires_reasoning` fields are not part of these derived task records.
|
| 48 |
|
| 49 |
-
[The source manifest](definitions/factual-source.json) records the exact publisher CSV revision and SHA-256, selection rule and expected counts. [The selection implementation](https://github.com/lexmount/webstone/blob/
|
| 50 |
|
| 51 |
## Data and schema
|
| 52 |
|
|
@@ -59,6 +59,8 @@ Publisher row order, question text and reference answer text are retained. Case
|
|
| 59 |
|
| 60 |
The viewer's `input` is the unchanged question. Parse `task` as JSON for `id`, `query`, `reference_answer`, `reference_urls`, `topic`, `answer_type`, provenance and derivation. Records have `family: web`, `category: factual` and `language: en`.
|
| 61 |
|
|
|
|
|
|
|
| 62 |
## Load and evaluate
|
| 63 |
|
| 64 |
Choose a full Hugging Face commit from this dataset's revision history and retain it with your evaluation:
|
|
@@ -78,7 +80,7 @@ query = task["query"]
|
|
| 78 |
answer, supporting_urls = task["reference_answer"], task["reference_urls"]
|
| 79 |
```
|
| 80 |
|
| 81 |
-
`definitions.tar.gz` contains the same definition bytes as `definitions/`. Webstone's [data lock](https://github.com/lexmount/webstone/blob/
|
| 82 |
|
| 83 |
Reference Hit@k tests whether a supplied supporting URL appears in the first k results; MRR@k measures the first such hit's reciprocal rank. Answer Match@k checks answer cues in returned text. Source references are not an exhaustive set of relevant pages, and lexical answer matches do not establish that a result correctly answers the question. Report these measures separately from graded relevance. Searchable copies of the dataset can also create contamination; Webstone exposes a separate contamination diagnostic while retaining the original results.
|
| 84 |
|
|
@@ -92,6 +94,6 @@ Cite both upstream works when using this derivative, and identify the selected d
|
|
| 92 |
>
|
| 93 |
> Wei et al. (2024). [Measuring Short-Form Factuality in Large Language Models](https://github.com/openai/simple-evals). OpenAI SimpleQA.
|
| 94 |
|
| 95 |
-
For the harness, cite [Webstone: Benchmarking Browsers and Web Tools for AI Agents](https://github.com/lexmount/webstone/blob/
|
| 96 |
|
| 97 |
[LexBench-Web-Search](https://huggingface.co/datasets/Lexmount/LexBench-Web-Search) contains the separate scenario-query suites. [LexBench-Web-Fetch](https://huggingface.co/datasets/Lexmount/LexBench-Web-Fetch) and [LexBench-WebMainBench](https://huggingface.co/datasets/Lexmount/LexBench-WebMainBench) evaluate fetching and extraction with their own task populations.
|
|
|
|
| 46 |
|
| 47 |
Publisher row order, question text and reference answer text are retained. Case IDs use `sqv-` followed by the publisher's `original_index`, preserving the link to OpenAI SimpleQA. Supporting URLs are split from the CSV field, deduplicated in order, and stripped of unmatched trailing closing parentheses. Topic and answer-type labels are retained. The publisher's `multi_step` and `requires_reasoning` fields are not part of these derived task records.
|
| 48 |
|
| 49 |
+
[The source manifest](definitions/factual-source.json) records the exact publisher CSV revision and SHA-256, selection rule and expected counts. [The selection implementation](https://github.com/lexmount/webstone/blob/c8c21bd8e2654107d602cd31a89b4719369f11ed/src/lexbench_websearch/simpleqa.py) rebuilds the population from that input. This release contains no search-provider rankings, relevance grades or measured results.
|
| 50 |
|
| 51 |
## Data and schema
|
| 52 |
|
|
|
|
| 59 |
|
| 60 |
The viewer's `input` is the unchanged question. Parse `task` as JSON for `id`, `query`, `reference_answer`, `reference_urls`, `topic`, `answer_type`, provenance and derivation. Records have `family: web`, `category: factual` and `language: en`.
|
| 61 |
|
| 62 |
+
`reference_urls` retains the publisher's URLs. `equivalent_reference_urls` adds independently verified copies of those documents. Candidates come from fixed mirror URL templates applied to the reference URLs, without consulting search results. A copy needs the same article title and at least 80% matching characters in its normalized lead text, in contiguous blocks of at least 12 characters. Verification uses up to 4,000 lead characters and requires at least 200 normalized characters from each page. Failed requests and short pages establish no equivalence. `equivalent_reference_provenance` records the reference URL, verification method, observation date and source hashes for each accepted copy. The dataset's `reference_equivalence` record fixes these rules; `url_equivalence` selects the generic document-URL policy.
|
| 63 |
+
|
| 64 |
## Load and evaluate
|
| 65 |
|
| 66 |
Choose a full Hugging Face commit from this dataset's revision history and retain it with your evaluation:
|
|
|
|
| 80 |
answer, supporting_urls = task["reference_answer"], task["reference_urls"]
|
| 81 |
```
|
| 82 |
|
| 83 |
+
`definitions.tar.gz` contains the same definition bytes as `definitions/`. Webstone's [data lock](https://github.com/lexmount/webstone/blob/c8c21bd8e2654107d602cd31a89b4719369f11ed/src/harness/benchmarks/data-lock.json) selects the installed snapshot. Follow the [Web Search guide](https://github.com/lexmount/webstone/blob/c8c21bd8e2654107d602cd31a89b4719369f11ed/src/harness/web_search/README.md) to declare a profile, collect provider responses and build a report with saved evidence. The certified collector is `scripts/run_certified_benchmark.py web-search`.
|
| 84 |
|
| 85 |
Reference Hit@k tests whether a supplied supporting URL appears in the first k results; MRR@k measures the first such hit's reciprocal rank. Answer Match@k checks answer cues in returned text. Source references are not an exhaustive set of relevant pages, and lexical answer matches do not establish that a result correctly answers the question. Report these measures separately from graded relevance. Searchable copies of the dataset can also create contamination; Webstone exposes a separate contamination diagnostic while retaining the original results.
|
| 86 |
|
|
|
|
| 94 |
>
|
| 95 |
> Wei et al. (2024). [Measuring Short-Form Factuality in Large Language Models](https://github.com/openai/simple-evals). OpenAI SimpleQA.
|
| 96 |
|
| 97 |
+
For the harness, cite [Webstone: Benchmarking Browsers and Web Tools for AI Agents](https://github.com/lexmount/webstone/blob/c8c21bd8e2654107d602cd31a89b4719369f11ed/CITATION.cff).
|
| 98 |
|
| 99 |
[LexBench-Web-Search](https://huggingface.co/datasets/Lexmount/LexBench-Web-Search) contains the separate scenario-query suites. [LexBench-Web-Fetch](https://huggingface.co/datasets/Lexmount/LexBench-Web-Fetch) and [LexBench-WebMainBench](https://huggingface.co/datasets/Lexmount/LexBench-WebMainBench) evaluate fetching and extraction with their own task populations.
|
data/factual-v1.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
definitions.tar.gz
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6f887fdf342a9c7c258d1737711027a754491f468ac705208ae7eab25c94c141
|
| 3 |
+
size 214869
|
definitions/datasets/factual-v1.json
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
definitions/reference-equivalence.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema": "simpleqa-reference-revision/1",
|
| 3 |
+
"root_cause": "Literal URL matching missed copies of the same article. Added URLs derive from publisher references and require matching observed titles and original lead text.",
|
| 4 |
+
"regression_ids": [
|
| 5 |
+
"generic-document-urls",
|
| 6 |
+
"reference-mirror-title-and-lead"
|
| 7 |
+
],
|
| 8 |
+
"provider_calls": 0,
|
| 9 |
+
"source_tooling_revision": "c8c21bd8e2654107d602cd31a89b4719369f11ed",
|
| 10 |
+
"source_dataset_revision": "bd2a330cf45b4f8ab77f8b9111c50c2faa01cc75",
|
| 11 |
+
"before_dataset_sha256": "b3422267e0be4a68b71e533356e14531b6aba1d2e57ccde962e6e3e00327d253",
|
| 12 |
+
"after_dataset_sha256": "41e36f3e5f8c954192b6ecbd6f8b5a6bc86cd1950c79eafb89d0a96c3aa4f536",
|
| 13 |
+
"policy": {
|
| 14 |
+
"method": "reference-title-lead/1",
|
| 15 |
+
"minimum_lead_characters": 200,
|
| 16 |
+
"maximum_lead_characters": 4000,
|
| 17 |
+
"minimum_block_overlap": 0.8,
|
| 18 |
+
"candidate_scope": "Publisher reference URLs only",
|
| 19 |
+
"url_policy": "document-url/2"
|
| 20 |
+
},
|
| 21 |
+
"evidence": {
|
| 22 |
+
"input_cases": 831,
|
| 23 |
+
"reference_url_count": 2451,
|
| 24 |
+
"observed_urls": 3209,
|
| 25 |
+
"available_pages": 1987,
|
| 26 |
+
"with_article_lead": 1981,
|
| 27 |
+
"qualified_aliases": 691,
|
| 28 |
+
"cases_with_aliases": 474,
|
| 29 |
+
"observed_capture_cost_usd": 1.987,
|
| 30 |
+
"transport_cost_unknown_batches": 1,
|
| 31 |
+
"observations_sha256": "c953bfda568ecdef261e2087caaf24104be8dc7fc246dcc485116be26921d2cb",
|
| 32 |
+
"extended_dataset_sha256": "41e36f3e5f8c954192b6ecbd6f8b5a6bc86cd1950c79eafb89d0a96c3aa4f536"
|
| 33 |
+
},
|
| 34 |
+
"capture_contracts_sha256": [
|
| 35 |
+
"1e4065df58e290843fd5b1831cdc5d48e3c80a7186c24a92247061cb683c8dce",
|
| 36 |
+
"4211ff68079835382a8a2a157a1bdba1e8635c0c2235018b8ac1262922489800"
|
| 37 |
+
]
|
| 38 |
+
}
|
definitions/registry.json
CHANGED
|
@@ -3,7 +3,7 @@
|
|
| 3 |
"factual-v1": {
|
| 4 |
"count": 831,
|
| 5 |
"path": "datasets/factual-v1.json",
|
| 6 |
-
"sha256": "
|
| 7 |
}
|
| 8 |
}
|
| 9 |
}
|
|
|
|
| 3 |
"factual-v1": {
|
| 4 |
"count": 831,
|
| 5 |
"path": "datasets/factual-v1.json",
|
| 6 |
+
"sha256": "41e36f3e5f8c954192b6ecbd6f8b5a6bc86cd1950c79eafb89d0a96c3aa4f536"
|
| 7 |
}
|
| 8 |
}
|
| 9 |
}
|
manifest.json
CHANGED
|
@@ -1,14 +1,14 @@
|
|
| 1 |
{
|
| 2 |
"archive": {
|
| 3 |
"path": "definitions.tar.gz",
|
| 4 |
-
"sha256": "
|
| 5 |
-
"size":
|
| 6 |
},
|
| 7 |
"datasets": {
|
| 8 |
"factual-v1": {
|
| 9 |
"count": 831,
|
| 10 |
"path": "datasets/factual-v1.json",
|
| 11 |
-
"sha256": "
|
| 12 |
}
|
| 13 |
},
|
| 14 |
"files": {
|
|
@@ -25,32 +25,36 @@
|
|
| 25 |
"size": 1817
|
| 26 |
},
|
| 27 |
"data/factual-v1.jsonl": {
|
| 28 |
-
"sha256": "
|
| 29 |
-
"size":
|
| 30 |
},
|
| 31 |
"definitions/NOTICE.txt": {
|
| 32 |
"sha256": "4787042ef9f5d9e353edd7c747ac15c0fe4a2a572f0dd3ce1f52cf8c6549215c",
|
| 33 |
"size": 3634
|
| 34 |
},
|
| 35 |
"definitions/datasets/factual-v1.json": {
|
| 36 |
-
"sha256": "
|
| 37 |
-
"size":
|
| 38 |
},
|
| 39 |
"definitions/factual-source.json": {
|
| 40 |
"sha256": "b2f0f48cece8fa8c6449acd387a2936c23cf678801138d48b70952d24a2210db",
|
| 41 |
"size": 1032
|
| 42 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
"definitions/registry.json": {
|
| 44 |
-
"sha256": "
|
| 45 |
"size": 195
|
| 46 |
}
|
| 47 |
},
|
| 48 |
-
"schema": "
|
| 49 |
-
"source_repository": "https://
|
| 50 |
-
"source_revision": "
|
| 51 |
"tooling": {
|
| 52 |
-
"repository": "https://github.com/lexmount/
|
| 53 |
-
"revision": "
|
| 54 |
},
|
| 55 |
"track": "simpleqa-verified",
|
| 56 |
"viewer": {
|
|
@@ -63,5 +67,9 @@
|
|
| 63 |
}
|
| 64 |
],
|
| 65 |
"language_annotations": {}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
}
|
| 67 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"archive": {
|
| 3 |
"path": "definitions.tar.gz",
|
| 4 |
+
"sha256": "6f887fdf342a9c7c258d1737711027a754491f468ac705208ae7eab25c94c141",
|
| 5 |
+
"size": 214869
|
| 6 |
},
|
| 7 |
"datasets": {
|
| 8 |
"factual-v1": {
|
| 9 |
"count": 831,
|
| 10 |
"path": "datasets/factual-v1.json",
|
| 11 |
+
"sha256": "41e36f3e5f8c954192b6ecbd6f8b5a6bc86cd1950c79eafb89d0a96c3aa4f536"
|
| 12 |
}
|
| 13 |
},
|
| 14 |
"files": {
|
|
|
|
| 25 |
"size": 1817
|
| 26 |
},
|
| 27 |
"data/factual-v1.jsonl": {
|
| 28 |
+
"sha256": "900c0f5fe6bd71e877076ca359de8edb2eb08f30669c990ceb93a3051a714dc5",
|
| 29 |
+
"size": 1353295
|
| 30 |
},
|
| 31 |
"definitions/NOTICE.txt": {
|
| 32 |
"sha256": "4787042ef9f5d9e353edd7c747ac15c0fe4a2a572f0dd3ce1f52cf8c6549215c",
|
| 33 |
"size": 3634
|
| 34 |
},
|
| 35 |
"definitions/datasets/factual-v1.json": {
|
| 36 |
+
"sha256": "41e36f3e5f8c954192b6ecbd6f8b5a6bc86cd1950c79eafb89d0a96c3aa4f536",
|
| 37 |
+
"size": 1302104
|
| 38 |
},
|
| 39 |
"definitions/factual-source.json": {
|
| 40 |
"sha256": "b2f0f48cece8fa8c6449acd387a2936c23cf678801138d48b70952d24a2210db",
|
| 41 |
"size": 1032
|
| 42 |
},
|
| 43 |
+
"definitions/reference-equivalence.json": {
|
| 44 |
+
"sha256": "04d01ec31c9b4f37ab8e52769a8f5a1161996fac3114f5aaae7d688019c074fe",
|
| 45 |
+
"size": 1617
|
| 46 |
+
},
|
| 47 |
"definitions/registry.json": {
|
| 48 |
+
"sha256": "73ce568f3aaace5002cc1f7bec29939408a20160c6173ce580eab60a50d56095",
|
| 49 |
"size": 195
|
| 50 |
}
|
| 51 |
},
|
| 52 |
+
"schema": "webstone-dataset-release/1",
|
| 53 |
+
"source_repository": "https://huggingface.co/datasets/Lexmount/LexBench-SimpleQA-Verified",
|
| 54 |
+
"source_revision": "6c5f51e9d70f5dc958b708b5e8a736646fa2132c",
|
| 55 |
"tooling": {
|
| 56 |
+
"repository": "https://github.com/lexmount/webstone",
|
| 57 |
+
"revision": "c8c21bd8e2654107d602cd31a89b4719369f11ed"
|
| 58 |
},
|
| 59 |
"track": "simpleqa-verified",
|
| 60 |
"viewer": {
|
|
|
|
| 67 |
}
|
| 68 |
],
|
| 69 |
"language_annotations": {}
|
| 70 |
+
},
|
| 71 |
+
"viewer_filters": {
|
| 72 |
+
"viewer_exclude_datasets": [],
|
| 73 |
+
"viewer_exclude_tracks": {}
|
| 74 |
}
|
| 75 |
}
|