Fix multi-config loading with uniform JSONL files
Browse filesAdd a deterministic tidy metric_rows JSONL mirror and point both Hugging Face configs to JSONL files. The approved benchmark values, nested evaluations, CSV download, CC BY 4.0 license, disclosure, and campaign backlink are unchanged. Adds regression coverage for config-format compatibility and exact CSV/JSONL equivalence.
- Makefile +2 -1
- README.md +80 -12
- data/metric_rows.jsonl +0 -0
- data/summary.json +43 -0
- scripts/build_dataset.py +165 -45
- tests/test_dataset.py +231 -12
Makefile
CHANGED
|
@@ -35,7 +35,8 @@ validate:
|
|
| 35 |
fi
|
| 36 |
|
| 37 |
hashes:
|
| 38 |
-
@sha256sum $(EVALS) $(DATA)/
|
|
|
|
| 39 |
|
| 40 |
help:
|
| 41 |
@sed -n '1,13p' Makefile
|
|
|
|
| 35 |
fi
|
| 36 |
|
| 37 |
hashes:
|
| 38 |
+
@sha256sum $(EVALS) $(DATA)/metric_rows.jsonl $(DATA)/benchmarks.csv \
|
| 39 |
+
$(DATA)/benchmarks.jsonl $(DATA)/summary.json README.md LICENSE
|
| 40 |
|
| 41 |
help:
|
| 42 |
@sed -n '1,13p' Makefile
|
README.md
CHANGED
|
@@ -19,7 +19,7 @@ configs:
|
|
| 19 |
- config_name: metric_rows
|
| 20 |
data_files:
|
| 21 |
- split: train
|
| 22 |
-
path: data/
|
| 23 |
- config_name: evaluations
|
| 24 |
data_files:
|
| 25 |
- split: train
|
|
@@ -52,19 +52,38 @@ unchanged — three scorecards in this snapshot sit below 2.0 out of 10, and two
|
|
| 52 |
|
| 53 |
## Dataset structure
|
| 54 |
|
| 55 |
-
Two views of the same 67 scorecards. Nothing is aggregated or re-scored
|
|
|
|
| 56 |
|
| 57 |
-
| Config | File |
|
| 58 |
| --- | --- | --- | --- |
|
| 59 |
-
| `metric_rows` | `data/
|
| 60 |
| `evaluations` | `data/benchmarks.jsonl` | 67 | one record per provider × capability, metrics nested |
|
| 61 |
|
| 62 |
-
|
| 63 |
-
|
|
|
|
|
|
|
| 64 |
|
| 65 |
-
|
| 66 |
|
| 67 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
| --- | --- | --- |
|
| 69 |
| `evaluation_id` | string | `{provider_id}:{capability_id}`, unique per scorecard |
|
| 70 |
| `provider_id` | string | Provider slug as published by the source |
|
|
@@ -90,9 +109,43 @@ hash, per-capability inventories). Every figure in it is derived by the generato
|
|
| 90 |
| `source_url` | url | `https://nativeport.ai/evals.json` |
|
| 91 |
| `snapshot_sha256` | string | SHA-256 of the exact source snapshot these rows came from |
|
| 92 |
|
| 93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
and verifiable on its own.
|
| 95 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
### `data/benchmarks.jsonl` — one record per evaluation
|
| 97 |
|
| 98 |
```json
|
|
@@ -196,13 +249,21 @@ rows = load_dataset("nativeport/web-access-api-benchmarks", "metric_rows", split
|
|
| 196 |
evals = load_dataset("nativeport/web-access-api-benchmarks", "evaluations", split="train")
|
| 197 |
```
|
| 198 |
|
|
|
|
|
|
|
|
|
|
| 199 |
Cheapest search provider that clears a recall floor — the typical routing question,
|
| 200 |
answered from the tidy view with pandas:
|
| 201 |
|
| 202 |
```python
|
| 203 |
import pandas as pd
|
| 204 |
|
| 205 |
-
df = pd.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 206 |
search = df[df.capability_id == "search"]
|
| 207 |
|
| 208 |
wide = search.pivot_table(
|
|
@@ -298,8 +359,10 @@ bytes produce byte-identical outputs. No wall-clock timestamp is written anywher
|
|
| 298 |
a rebuild can be diffed directly against the published files. Numeric values are
|
| 299 |
carried across as exact source tokens — the generator verifies that each emitted
|
| 300 |
number serialises character-for-character back to the token it read, and aborts rather
|
| 301 |
-
than emit a rounded stand-in.
|
| 302 |
-
files.
|
|
|
|
|
|
|
| 303 |
|
| 304 |
Fields present in the source but **excluded by design**: gateway routing and
|
| 305 |
authentication strings, list prices, latency prose, marketing summaries,
|
|
@@ -342,6 +405,11 @@ measurements and the provenance needed to interpret them are published here.
|
|
| 342 |
numbers are not silently rewritten in place.
|
| 343 |
- Schema changes that are not backward compatible will be described in the commit that
|
| 344 |
makes them and reflected in the tables above.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 345 |
- No update cadence is promised here. `measured_date` and `run_id` are on every row
|
| 346 |
precisely so a consumer can decide for itself whether the snapshot is still fresh
|
| 347 |
enough to act on.
|
|
|
|
| 19 |
- config_name: metric_rows
|
| 20 |
data_files:
|
| 21 |
- split: train
|
| 22 |
+
path: data/metric_rows.jsonl
|
| 23 |
- config_name: evaluations
|
| 24 |
data_files:
|
| 25 |
- split: train
|
|
|
|
| 52 |
|
| 53 |
## Dataset structure
|
| 54 |
|
| 55 |
+
Two configs, two views of the same 67 scorecards. Nothing is aggregated or re-scored
|
| 56 |
+
between them.
|
| 57 |
|
| 58 |
+
| Config | File | Records | Grain |
|
| 59 |
| --- | --- | --- | --- |
|
| 60 |
+
| `metric_rows` | `data/metric_rows.jsonl` | 297 | one record per provider × capability × metric (tidy/long) |
|
| 61 |
| `evaluations` | `data/benchmarks.jsonl` | 67 | one record per provider × capability, metrics nested |
|
| 62 |
|
| 63 |
+
Both configs are JSONL deliberately. The Hub resolves a single packaged loader for the
|
| 64 |
+
whole repository from the declared config data files and then applies it to every
|
| 65 |
+
config, so a repository that mixes formats across its configs ends up parsing one of
|
| 66 |
+
them with the other's reader.
|
| 67 |
|
| 68 |
+
Two more files ship alongside them and back no config:
|
| 69 |
|
| 70 |
+
| File | Rows | What it is |
|
| 71 |
+
| --- | --- | --- |
|
| 72 |
+
| `data/benchmarks.csv` | 297 | the tidy view as CSV — same rows, same field order, same values as `data/metric_rows.jsonl` |
|
| 73 |
+
| `data/summary.json` | — | computed snapshot facts: counts, date range, source hash, per-capability and per-metric inventories, artifact digests |
|
| 74 |
+
|
| 75 |
+
`data/benchmarks.csv` is a download rather than a config. `load_dataset` never reads it,
|
| 76 |
+
but `pd.read_csv("hf://datasets/nativeport/web-access-api-benchmarks/data/benchmarks.csv")`,
|
| 77 |
+
`huggingface_hub.hf_hub_download` and a plain browser download all do. Both tidy files
|
| 78 |
+
are written in one pass from one row builder, so they cannot drift; the test suite
|
| 79 |
+
compares them field by field across all 297 rows, including the exact text of every
|
| 80 |
+
number. Every figure in `data/summary.json` is derived by the generator, never typed.
|
| 81 |
+
|
| 82 |
+
### `data/metric_rows.jsonl` and `data/benchmarks.csv` — tidy metric rows
|
| 83 |
+
|
| 84 |
+
Both carry these 23 fields, in this order:
|
| 85 |
+
|
| 86 |
+
| Field | Type | Description |
|
| 87 |
| --- | --- | --- |
|
| 88 |
| `evaluation_id` | string | `{provider_id}:{capability_id}`, unique per scorecard |
|
| 89 |
| `provider_id` | string | Provider slug as published by the source |
|
|
|
|
| 109 |
| `source_url` | url | `https://nativeport.ai/evals.json` |
|
| 110 |
| `snapshot_sha256` | string | SHA-256 of the exact source snapshot these rows came from |
|
| 111 |
|
| 112 |
+
In `data/metric_rows.jsonl` the types above are the JSON types: `metric_raw` and
|
| 113 |
+
`composite_score` are JSON numbers, `metric_index` / `rank` / `rank_of` /
|
| 114 |
+
`composite_scale_max` are integers, and `is_capability_top` is a boolean. CSV has no
|
| 115 |
+
types, so `data/benchmarks.csv` carries the same values as text — numbers written as
|
| 116 |
+
the exact source token, booleans as `true` / `false`.
|
| 117 |
+
|
| 118 |
+
Provenance fields repeat on every row so that a filtered slice stays interpretable
|
| 119 |
and verifiable on its own.
|
| 120 |
|
| 121 |
+
```json
|
| 122 |
+
{
|
| 123 |
+
"evaluation_id": "serper:serp",
|
| 124 |
+
"provider_id": "serper",
|
| 125 |
+
"provider_name": "Serper",
|
| 126 |
+
"provider_group": "Search",
|
| 127 |
+
"provider_category": "Google SERP scrape",
|
| 128 |
+
"capability_id": "serp",
|
| 129 |
+
"capability_label": "SERP verticals",
|
| 130 |
+
"metric_key": "latency_p50_ms",
|
| 131 |
+
"metric_label": "Latency p50",
|
| 132 |
+
"metric_raw": 886.0,
|
| 133 |
+
"metric_display": "886 ms",
|
| 134 |
+
"metric_index": 1,
|
| 135 |
+
"composite_score": 9.13,
|
| 136 |
+
"composite_scale_max": 10,
|
| 137 |
+
"rank": 1,
|
| 138 |
+
"rank_of": 4,
|
| 139 |
+
"is_capability_top": true,
|
| 140 |
+
"measured_date": "2026-08-05",
|
| 141 |
+
"note": "Lowest cost and latency; Google only.",
|
| 142 |
+
"provider_page_url": "https://nativeport.ai/providers/serper/",
|
| 143 |
+
"run_id": "2026-08-05",
|
| 144 |
+
"source_url": "https://nativeport.ai/evals.json",
|
| 145 |
+
"snapshot_sha256": "f46bf416803d5adf696c66ac14a6cbf06f4dfa39727cca164e8f3872abd6ed9b"
|
| 146 |
+
}
|
| 147 |
+
```
|
| 148 |
+
|
| 149 |
### `data/benchmarks.jsonl` — one record per evaluation
|
| 150 |
|
| 151 |
```json
|
|
|
|
| 249 |
evals = load_dataset("nativeport/web-access-api-benchmarks", "evaluations", split="train")
|
| 250 |
```
|
| 251 |
|
| 252 |
+
Both configs load as JSON Lines. `rows` has the 23 flat fields listed above, `evals` the
|
| 253 |
+
nested `metrics` list.
|
| 254 |
+
|
| 255 |
Cheapest search provider that clears a recall floor — the typical routing question,
|
| 256 |
answered from the tidy view with pandas:
|
| 257 |
|
| 258 |
```python
|
| 259 |
import pandas as pd
|
| 260 |
|
| 261 |
+
df = pd.read_json(
|
| 262 |
+
"hf://datasets/nativeport/web-access-api-benchmarks/data/metric_rows.jsonl", lines=True
|
| 263 |
+
)
|
| 264 |
+
# The CSV mirror gives the same frame, for tools that prefer a spreadsheet:
|
| 265 |
+
# df = pd.read_csv("hf://datasets/nativeport/web-access-api-benchmarks/data/benchmarks.csv")
|
| 266 |
+
|
| 267 |
search = df[df.capability_id == "search"]
|
| 268 |
|
| 269 |
wide = search.pivot_table(
|
|
|
|
| 359 |
a rebuild can be diffed directly against the published files. Numeric values are
|
| 360 |
carried across as exact source tokens — the generator verifies that each emitted
|
| 361 |
number serialises character-for-character back to the token it read, and aborts rather
|
| 362 |
+
than emit a rounded stand-in. That same serialiser writes the CSV cell and the JSON
|
| 363 |
+
number, which is why the two tidy files agree token for token. `data/summary.json`
|
| 364 |
+
records the SHA-256 of all three data files, together with the config-to-file mapping
|
| 365 |
+
the front matter declares.
|
| 366 |
|
| 367 |
Fields present in the source but **excluded by design**: gateway routing and
|
| 368 |
authentication strings, list prices, latency prose, marketing summaries,
|
|
|
|
| 405 |
numbers are not silently rewritten in place.
|
| 406 |
- Schema changes that are not backward compatible will be described in the commit that
|
| 407 |
makes them and reflected in the tables above.
|
| 408 |
+
- File layout, for anyone who loaded an earlier revision: the `metric_rows` config is
|
| 409 |
+
backed by `data/metric_rows.jsonl`. It previously pointed at `data/benchmarks.csv`,
|
| 410 |
+
which left the two configs in different formats and made the Hub read one of them
|
| 411 |
+
with the wrong parser. No row, field or measured value changed — only the file the
|
| 412 |
+
config resolves to — and `data/benchmarks.csv` still ships, unchanged, as a download.
|
| 413 |
- No update cadence is promised here. `measured_date` and `run_id` are on every row
|
| 414 |
precisely so a consumer can decide for itself whether the snapshot is still fresh
|
| 415 |
enough to act on.
|
data/metric_rows.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/summary.json
CHANGED
|
@@ -876,7 +876,49 @@
|
|
| 876 |
]
|
| 877 |
}
|
| 878 |
],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 879 |
"outputs": {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 880 |
"benchmarks.csv": {
|
| 881 |
"kind": "csv",
|
| 882 |
"data_rows": 297,
|
|
@@ -906,6 +948,7 @@
|
|
| 906 |
"source_url",
|
| 907 |
"snapshot_sha256"
|
| 908 |
],
|
|
|
|
| 909 |
"bytes": 101064,
|
| 910 |
"sha256": "6153966709823465ffa1c369b4e48170d793bbece92a25938dd63fec9bd5ceb9"
|
| 911 |
},
|
|
|
|
| 876 |
]
|
| 877 |
}
|
| 878 |
],
|
| 879 |
+
"hub_config_data_files": {
|
| 880 |
+
"metric_rows": "data/metric_rows.jsonl",
|
| 881 |
+
"evaluations": "data/benchmarks.jsonl"
|
| 882 |
+
},
|
| 883 |
+
"hub_config_data_file_format": [
|
| 884 |
+
"jsonl"
|
| 885 |
+
],
|
| 886 |
+
"downloadable_only_artifacts": [
|
| 887 |
+
"data/benchmarks.csv"
|
| 888 |
+
],
|
| 889 |
"outputs": {
|
| 890 |
+
"metric_rows.jsonl": {
|
| 891 |
+
"kind": "jsonl",
|
| 892 |
+
"records": 297,
|
| 893 |
+
"fields": 23,
|
| 894 |
+
"field_names": [
|
| 895 |
+
"evaluation_id",
|
| 896 |
+
"provider_id",
|
| 897 |
+
"provider_name",
|
| 898 |
+
"provider_group",
|
| 899 |
+
"provider_category",
|
| 900 |
+
"capability_id",
|
| 901 |
+
"capability_label",
|
| 902 |
+
"metric_key",
|
| 903 |
+
"metric_label",
|
| 904 |
+
"metric_raw",
|
| 905 |
+
"metric_display",
|
| 906 |
+
"metric_index",
|
| 907 |
+
"composite_score",
|
| 908 |
+
"composite_scale_max",
|
| 909 |
+
"rank",
|
| 910 |
+
"rank_of",
|
| 911 |
+
"is_capability_top",
|
| 912 |
+
"measured_date",
|
| 913 |
+
"note",
|
| 914 |
+
"provider_page_url",
|
| 915 |
+
"run_id",
|
| 916 |
+
"source_url",
|
| 917 |
+
"snapshot_sha256"
|
| 918 |
+
],
|
| 919 |
+
"bytes": 228423,
|
| 920 |
+
"sha256": "0f9f2a8c395997579fce91123e9c0996a7d164c91a581acbef5bc956739aacda"
|
| 921 |
+
},
|
| 922 |
"benchmarks.csv": {
|
| 923 |
"kind": "csv",
|
| 924 |
"data_rows": 297,
|
|
|
|
| 948 |
"source_url",
|
| 949 |
"snapshot_sha256"
|
| 950 |
],
|
| 951 |
+
"mirrors": "metric_rows.jsonl",
|
| 952 |
"bytes": 101064,
|
| 953 |
"sha256": "6153966709823465ffa1c369b4e48170d793bbece92a25938dd63fec9bd5ceb9"
|
| 954 |
},
|
scripts/build_dataset.py
CHANGED
|
@@ -1,12 +1,22 @@
|
|
| 1 |
#!/usr/bin/env python3
|
| 2 |
"""Deterministic generator for the NativePort Web-Access API Benchmarks dataset.
|
| 3 |
|
| 4 |
-
Reads a NativePort ``evals.json`` snapshot and writes
|
| 5 |
|
| 6 |
-
data/
|
| 7 |
-
data/benchmarks.
|
|
|
|
| 8 |
data/summary.json computed snapshot facts (every count is derived, never typed)
|
| 9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
Design rules
|
| 11 |
------------
|
| 12 |
* Standard library only.
|
|
@@ -48,7 +58,29 @@ LEADERBOARDS_URL = "https://nativeport.ai/leaderboards/"
|
|
| 48 |
# has to assume the scale.
|
| 49 |
COMPOSITE_SCALE_MAX = 10
|
| 50 |
|
| 51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
"evaluation_id",
|
| 53 |
"provider_id",
|
| 54 |
"provider_name",
|
|
@@ -309,41 +341,91 @@ def extract_evaluations(document, snapshot_sha256):
|
|
| 309 |
# writers
|
| 310 |
# --------------------------------------------------------------------------- #
|
| 311 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 312 |
def write_csv(evaluations, path):
|
| 313 |
rows = 0
|
| 314 |
with path.open("w", encoding="utf-8", newline="") as handle:
|
| 315 |
writer = csv.writer(handle, lineterminator="\n", quoting=csv.QUOTE_MINIMAL)
|
| 316 |
-
writer.writerow(
|
| 317 |
-
for
|
| 318 |
-
for
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
evaluation["rank_of_text"],
|
| 337 |
-
"true" if evaluation["is_capability_top"] else "false",
|
| 338 |
-
evaluation["measured_date"],
|
| 339 |
-
evaluation["note"],
|
| 340 |
-
evaluation["provider_page_url"],
|
| 341 |
-
evaluation["run_id"],
|
| 342 |
-
evaluation["source_url"],
|
| 343 |
-
evaluation["snapshot_sha256"],
|
| 344 |
-
]
|
| 345 |
-
)
|
| 346 |
-
rows += 1
|
| 347 |
return rows
|
| 348 |
|
| 349 |
|
|
@@ -396,8 +478,14 @@ def sha256_of(path):
|
|
| 396 |
return hashlib.sha256(path.read_bytes()).hexdigest()
|
| 397 |
|
| 398 |
|
| 399 |
-
def build_summary(evaluations, meta,
|
| 400 |
-
"""Every figure here is computed from the extracted rows.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 401 |
provider_ids = sorted({e["provider_id"] for e in evaluations})
|
| 402 |
capability_ids = sorted({e["capability_id"] for e in evaluations})
|
| 403 |
measured_dates = sorted({e["measured_date"] for e in evaluations})
|
|
@@ -505,12 +593,26 @@ def build_summary(evaluations, meta, csv_path, jsonl_path, csv_rows, jsonl_recor
|
|
| 505 |
"capabilities": capabilities,
|
| 506 |
"providers": providers,
|
| 507 |
"metric_keys": metric_keys,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 508 |
"outputs": {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 509 |
csv_path.name: {
|
| 510 |
"kind": "csv",
|
| 511 |
"data_rows": csv_rows,
|
| 512 |
-
"columns": len(
|
| 513 |
-
"column_names": list(
|
|
|
|
| 514 |
"bytes": csv_path.stat().st_size,
|
| 515 |
"sha256": sha256_of(csv_path),
|
| 516 |
},
|
|
@@ -543,14 +645,27 @@ def build(input_path, output_dir):
|
|
| 543 |
evaluations, meta = extract_evaluations(document, snapshot_sha256)
|
| 544 |
require(evaluations, "no evaluations found in the snapshot")
|
| 545 |
|
| 546 |
-
|
| 547 |
-
|
| 548 |
-
|
|
|
|
| 549 |
|
|
|
|
| 550 |
csv_rows = write_csv(evaluations, csv_path)
|
| 551 |
jsonl_records = write_jsonl(evaluations, jsonl_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 552 |
summary = build_summary(
|
| 553 |
-
evaluations,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 554 |
)
|
| 555 |
write_summary(summary, summary_path)
|
| 556 |
return summary
|
|
@@ -572,7 +687,10 @@ def main(argv=None):
|
|
| 572 |
parser.add_argument(
|
| 573 |
"--output-dir",
|
| 574 |
default=str(default_output_dir()),
|
| 575 |
-
help=
|
|
|
|
|
|
|
|
|
|
| 576 |
)
|
| 577 |
args = parser.parse_args(argv)
|
| 578 |
|
|
@@ -598,7 +716,9 @@ def main(argv=None):
|
|
| 598 |
entry = outputs[name]
|
| 599 |
size = entry.get("data_rows", entry.get("records"))
|
| 600 |
print(f"wrote {name} ({size} rows, {entry['bytes']} bytes) {entry['sha256']}")
|
| 601 |
-
print("wrote
|
|
|
|
|
|
|
| 602 |
return 0
|
| 603 |
|
| 604 |
|
|
|
|
| 1 |
#!/usr/bin/env python3
|
| 2 |
"""Deterministic generator for the NativePort Web-Access API Benchmarks dataset.
|
| 3 |
|
| 4 |
+
Reads a NativePort ``evals.json`` snapshot and writes four artifacts:
|
| 5 |
|
| 6 |
+
data/metric_rows.jsonl tidy/long form, one record per provider x capability x metric
|
| 7 |
+
data/benchmarks.csv the same tidy rows as CSV, value for value identical
|
| 8 |
+
data/benchmarks.jsonl one record per provider x capability evaluation, metrics nested
|
| 9 |
data/summary.json computed snapshot facts (every count is derived, never typed)
|
| 10 |
|
| 11 |
+
Both Hugging Face configs declared in README.md front matter are backed by JSONL
|
| 12 |
+
(``metric_rows`` -> data/metric_rows.jsonl, ``evaluations`` -> data/benchmarks.jsonl).
|
| 13 |
+
The Hub resolves a single packaged builder for a whole repository from the declared
|
| 14 |
+
config data files and applies it to every config, so a repository that mixes CSV and
|
| 15 |
+
JSONL across configs has the wrong parser applied to one of them.
|
| 16 |
+
``data/benchmarks.csv`` therefore stays a plain downloadable artifact and backs no
|
| 17 |
+
config. The tidy CSV and the tidy JSONL are emitted from one row builder, so the two
|
| 18 |
+
cannot drift.
|
| 19 |
+
|
| 20 |
Design rules
|
| 21 |
------------
|
| 22 |
* Standard library only.
|
|
|
|
| 58 |
# has to assume the scale.
|
| 59 |
COMPOSITE_SCALE_MAX = 10
|
| 60 |
|
| 61 |
+
# Output file names. The tidy view is published twice from one row builder: as JSONL,
|
| 62 |
+
# which backs the `metric_rows` config, and as CSV, which is a downloadable artifact
|
| 63 |
+
# only. See HUB_CONFIG_DATA_FILES below for why the config cannot point at the CSV.
|
| 64 |
+
METRIC_ROWS_JSONL_NAME = "metric_rows.jsonl"
|
| 65 |
+
CSV_NAME = "benchmarks.csv"
|
| 66 |
+
EVALUATIONS_JSONL_NAME = "benchmarks.jsonl"
|
| 67 |
+
SUMMARY_NAME = "summary.json"
|
| 68 |
+
|
| 69 |
+
# The Hugging Face configs declared in README.md front matter, and the file backing
|
| 70 |
+
# each one. Both must be JSONL. The Hub resolves one packaged builder for the entire
|
| 71 |
+
# repository from the declared config data files and applies it to every config; mixing
|
| 72 |
+
# formats therefore parses one config with the other's reader. `data/benchmarks.csv` is
|
| 73 |
+
# deliberately absent from this mapping: it ships as a documented download, not as
|
| 74 |
+
# config data.
|
| 75 |
+
HUB_CONFIG_DATA_FILES = {
|
| 76 |
+
"metric_rows": f"data/{METRIC_ROWS_JSONL_NAME}",
|
| 77 |
+
"evaluations": f"data/{EVALUATIONS_JSONL_NAME}",
|
| 78 |
+
}
|
| 79 |
+
DOWNLOADABLE_ONLY_ARTIFACTS = [f"data/{CSV_NAME}"]
|
| 80 |
+
|
| 81 |
+
# Field order of the tidy view. The CSV header and the keys of every record in
|
| 82 |
+
# data/metric_rows.jsonl are this list, in this order.
|
| 83 |
+
TIDY_FIELDS = [
|
| 84 |
"evaluation_id",
|
| 85 |
"provider_id",
|
| 86 |
"provider_name",
|
|
|
|
| 341 |
# writers
|
| 342 |
# --------------------------------------------------------------------------- #
|
| 343 |
|
| 344 |
+
def tidy_record(evaluation, metric):
|
| 345 |
+
"""One tidy metric row, typed, with keys in ``TIDY_FIELDS`` order.
|
| 346 |
+
|
| 347 |
+
This is the single definition of the tidy view. ``write_metric_rows_jsonl`` writes
|
| 348 |
+
these values as JSON; ``write_csv`` writes the same values rendered as text by
|
| 349 |
+
``csv_cell``. Neither view can gain, lose or reorder a field without the other.
|
| 350 |
+
"""
|
| 351 |
+
record = {
|
| 352 |
+
"evaluation_id": evaluation["evaluation_id"],
|
| 353 |
+
"provider_id": evaluation["provider_id"],
|
| 354 |
+
"provider_name": evaluation["provider_name"],
|
| 355 |
+
"provider_group": evaluation["provider_group"],
|
| 356 |
+
"provider_category": evaluation["provider_category"],
|
| 357 |
+
"capability_id": evaluation["capability_id"],
|
| 358 |
+
"capability_label": evaluation["capability_label"],
|
| 359 |
+
"metric_key": metric["metric_key"],
|
| 360 |
+
"metric_label": metric["metric_label"],
|
| 361 |
+
"metric_raw": metric["raw_value"],
|
| 362 |
+
"metric_display": metric["display_value"],
|
| 363 |
+
"metric_index": metric["metric_index"],
|
| 364 |
+
"composite_score": evaluation["composite_score"],
|
| 365 |
+
"composite_scale_max": evaluation["composite_scale_max"],
|
| 366 |
+
"rank": evaluation["rank"],
|
| 367 |
+
"rank_of": evaluation["rank_of"],
|
| 368 |
+
"is_capability_top": evaluation["is_capability_top"],
|
| 369 |
+
"measured_date": evaluation["measured_date"],
|
| 370 |
+
"note": evaluation["note"],
|
| 371 |
+
"provider_page_url": evaluation["provider_page_url"],
|
| 372 |
+
"run_id": evaluation["run_id"],
|
| 373 |
+
"source_url": evaluation["source_url"],
|
| 374 |
+
"snapshot_sha256": evaluation["snapshot_sha256"],
|
| 375 |
+
}
|
| 376 |
+
require(
|
| 377 |
+
list(record) == TIDY_FIELDS,
|
| 378 |
+
"tidy record fields drifted from TIDY_FIELDS: "
|
| 379 |
+
f"{[f for f in record if f not in TIDY_FIELDS]} / "
|
| 380 |
+
f"{[f for f in TIDY_FIELDS if f not in record]}",
|
| 381 |
+
)
|
| 382 |
+
return record
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
def tidy_records(evaluations):
|
| 386 |
+
for evaluation in evaluations:
|
| 387 |
+
for metric in evaluation["metrics"]:
|
| 388 |
+
yield tidy_record(evaluation, metric)
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
def csv_cell(value):
|
| 392 |
+
"""Render one tidy value as CSV text.
|
| 393 |
+
|
| 394 |
+
Numbers go through ``json_number_text``, the same serialiser the JSONL writer uses,
|
| 395 |
+
so a measurement reads identically in both files. ``exact_number`` has already
|
| 396 |
+
proved that this text is the source token character for character.
|
| 397 |
+
"""
|
| 398 |
+
if isinstance(value, bool):
|
| 399 |
+
return "true" if value else "false"
|
| 400 |
+
if isinstance(value, (int, float)):
|
| 401 |
+
return json_number_text(value)
|
| 402 |
+
return value
|
| 403 |
+
|
| 404 |
+
|
| 405 |
def write_csv(evaluations, path):
|
| 406 |
rows = 0
|
| 407 |
with path.open("w", encoding="utf-8", newline="") as handle:
|
| 408 |
writer = csv.writer(handle, lineterminator="\n", quoting=csv.QUOTE_MINIMAL)
|
| 409 |
+
writer.writerow(TIDY_FIELDS)
|
| 410 |
+
for record in tidy_records(evaluations):
|
| 411 |
+
writer.writerow([csv_cell(value) for value in record.values()])
|
| 412 |
+
rows += 1
|
| 413 |
+
return rows
|
| 414 |
+
|
| 415 |
+
|
| 416 |
+
def write_metric_rows_jsonl(evaluations, path):
|
| 417 |
+
"""The tidy view as JSONL, one record per metric row.
|
| 418 |
+
|
| 419 |
+
This is what the ``metric_rows`` Hugging Face config loads. It carries the same
|
| 420 |
+
fields as the CSV in the same order, with numbers as JSON numbers and
|
| 421 |
+
``is_capability_top`` as a JSON boolean.
|
| 422 |
+
"""
|
| 423 |
+
rows = 0
|
| 424 |
+
with path.open("w", encoding="utf-8", newline="") as handle:
|
| 425 |
+
for record in tidy_records(evaluations):
|
| 426 |
+
handle.write(json.dumps(record, ensure_ascii=False))
|
| 427 |
+
handle.write("\n")
|
| 428 |
+
rows += 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 429 |
return rows
|
| 430 |
|
| 431 |
|
|
|
|
| 478 |
return hashlib.sha256(path.read_bytes()).hexdigest()
|
| 479 |
|
| 480 |
|
| 481 |
+
def build_summary(evaluations, meta, artifacts, source_bytes):
|
| 482 |
+
"""Every figure here is computed from the extracted rows.
|
| 483 |
+
|
| 484 |
+
``artifacts`` maps each written data file to (path, record count).
|
| 485 |
+
"""
|
| 486 |
+
(metric_rows_path, metric_rows) = artifacts["metric_rows_jsonl"]
|
| 487 |
+
(csv_path, csv_rows) = artifacts["csv"]
|
| 488 |
+
(jsonl_path, jsonl_records) = artifacts["evaluations_jsonl"]
|
| 489 |
provider_ids = sorted({e["provider_id"] for e in evaluations})
|
| 490 |
capability_ids = sorted({e["capability_id"] for e in evaluations})
|
| 491 |
measured_dates = sorted({e["measured_date"] for e in evaluations})
|
|
|
|
| 593 |
"capabilities": capabilities,
|
| 594 |
"providers": providers,
|
| 595 |
"metric_keys": metric_keys,
|
| 596 |
+
"hub_config_data_files": dict(HUB_CONFIG_DATA_FILES),
|
| 597 |
+
"hub_config_data_file_format": sorted(
|
| 598 |
+
{Path(relative).suffix.lstrip(".") for relative in HUB_CONFIG_DATA_FILES.values()}
|
| 599 |
+
),
|
| 600 |
+
"downloadable_only_artifacts": list(DOWNLOADABLE_ONLY_ARTIFACTS),
|
| 601 |
"outputs": {
|
| 602 |
+
metric_rows_path.name: {
|
| 603 |
+
"kind": "jsonl",
|
| 604 |
+
"records": metric_rows,
|
| 605 |
+
"fields": len(TIDY_FIELDS),
|
| 606 |
+
"field_names": list(TIDY_FIELDS),
|
| 607 |
+
"bytes": metric_rows_path.stat().st_size,
|
| 608 |
+
"sha256": sha256_of(metric_rows_path),
|
| 609 |
+
},
|
| 610 |
csv_path.name: {
|
| 611 |
"kind": "csv",
|
| 612 |
"data_rows": csv_rows,
|
| 613 |
+
"columns": len(TIDY_FIELDS),
|
| 614 |
+
"column_names": list(TIDY_FIELDS),
|
| 615 |
+
"mirrors": metric_rows_path.name,
|
| 616 |
"bytes": csv_path.stat().st_size,
|
| 617 |
"sha256": sha256_of(csv_path),
|
| 618 |
},
|
|
|
|
| 645 |
evaluations, meta = extract_evaluations(document, snapshot_sha256)
|
| 646 |
require(evaluations, "no evaluations found in the snapshot")
|
| 647 |
|
| 648 |
+
metric_rows_path = output_dir / METRIC_ROWS_JSONL_NAME
|
| 649 |
+
csv_path = output_dir / CSV_NAME
|
| 650 |
+
jsonl_path = output_dir / EVALUATIONS_JSONL_NAME
|
| 651 |
+
summary_path = output_dir / SUMMARY_NAME
|
| 652 |
|
| 653 |
+
metric_rows = write_metric_rows_jsonl(evaluations, metric_rows_path)
|
| 654 |
csv_rows = write_csv(evaluations, csv_path)
|
| 655 |
jsonl_records = write_jsonl(evaluations, jsonl_path)
|
| 656 |
+
require(
|
| 657 |
+
metric_rows == csv_rows,
|
| 658 |
+
f"tidy views disagree: {metric_rows} JSONL records vs {csv_rows} CSV rows",
|
| 659 |
+
)
|
| 660 |
summary = build_summary(
|
| 661 |
+
evaluations,
|
| 662 |
+
meta,
|
| 663 |
+
{
|
| 664 |
+
"metric_rows_jsonl": (metric_rows_path, metric_rows),
|
| 665 |
+
"csv": (csv_path, csv_rows),
|
| 666 |
+
"evaluations_jsonl": (jsonl_path, jsonl_records),
|
| 667 |
+
},
|
| 668 |
+
source_bytes,
|
| 669 |
)
|
| 670 |
write_summary(summary, summary_path)
|
| 671 |
return summary
|
|
|
|
| 687 |
parser.add_argument(
|
| 688 |
"--output-dir",
|
| 689 |
default=str(default_output_dir()),
|
| 690 |
+
help=(
|
| 691 |
+
f"Directory to write {METRIC_ROWS_JSONL_NAME}, {CSV_NAME}, "
|
| 692 |
+
f"{EVALUATIONS_JSONL_NAME} and {SUMMARY_NAME} into."
|
| 693 |
+
),
|
| 694 |
)
|
| 695 |
args = parser.parse_args(argv)
|
| 696 |
|
|
|
|
| 716 |
entry = outputs[name]
|
| 717 |
size = entry.get("data_rows", entry.get("records"))
|
| 718 |
print(f"wrote {name} ({size} rows, {entry['bytes']} bytes) {entry['sha256']}")
|
| 719 |
+
print(f"wrote {SUMMARY_NAME}")
|
| 720 |
+
for config_name in sorted(summary["hub_config_data_files"]):
|
| 721 |
+
print(f"hub config {config_name} -> {summary['hub_config_data_files'][config_name]}")
|
| 722 |
return 0
|
| 723 |
|
| 724 |
|
tests/test_dataset.py
CHANGED
|
@@ -27,8 +27,20 @@ from pathlib import Path
|
|
| 27 |
ROOT = Path(__file__).resolve().parent.parent
|
| 28 |
DATA = ROOT / "data"
|
| 29 |
CSV_PATH = DATA / "benchmarks.csv"
|
|
|
|
| 30 |
JSONL_PATH = DATA / "benchmarks.jsonl"
|
| 31 |
SUMMARY_PATH = DATA / "summary.json"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
README_PATH = ROOT / "README.md"
|
| 33 |
LICENSE_PATH = ROOT / "LICENSE"
|
| 34 |
MAKEFILE_PATH = ROOT / "Makefile"
|
|
@@ -63,6 +75,7 @@ CC_BY_4_0_NORMALISED_SHA256 = (
|
|
| 63 |
PUBLIC_UPLOAD_SET = (
|
| 64 |
"README.md",
|
| 65 |
"LICENSE",
|
|
|
|
| 66 |
"data/benchmarks.csv",
|
| 67 |
"data/benchmarks.jsonl",
|
| 68 |
"data/summary.json",
|
|
@@ -103,6 +116,20 @@ def read_jsonl_records():
|
|
| 103 |
return [json.loads(line) for line in handle]
|
| 104 |
|
| 105 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
def read_summary():
|
| 107 |
return json.loads(SUMMARY_PATH.read_text(encoding="utf-8"))
|
| 108 |
|
|
@@ -230,7 +257,7 @@ class GeneratorDeterminismTest(unittest.TestCase):
|
|
| 230 |
digests.append(
|
| 231 |
tuple(
|
| 232 |
sha256_of(out / name)
|
| 233 |
-
for name in
|
| 234 |
)
|
| 235 |
)
|
| 236 |
self.assertEqual(digests[0], digests[1])
|
|
@@ -281,7 +308,7 @@ class GeneratorDeterminismTest(unittest.TestCase):
|
|
| 281 |
with tempfile.TemporaryDirectory() as workspace:
|
| 282 |
out = Path(workspace) / "out"
|
| 283 |
build_dataset.build(SOURCE_PATH, out)
|
| 284 |
-
for name in
|
| 285 |
self.assertEqual(
|
| 286 |
sha256_of(out / name),
|
| 287 |
sha256_of(DATA / name),
|
|
@@ -365,10 +392,11 @@ class SchemaAndCountConsistencyTest(unittest.TestCase):
|
|
| 365 |
def test_csv_header_matches_the_generator_contract(self):
|
| 366 |
with CSV_PATH.open(encoding="utf-8", newline="") as handle:
|
| 367 |
header = next(csv.reader(handle))
|
| 368 |
-
self.assertEqual(header, build_dataset.
|
| 369 |
|
| 370 |
def test_row_counts_agree_across_artifacts(self):
|
| 371 |
self.assertEqual(len(self.rows), self.summary["metric_row_count"])
|
|
|
|
| 372 |
self.assertEqual(len(self.records), self.summary["evaluation_count"])
|
| 373 |
self.assertEqual(
|
| 374 |
sum(len(record["metrics"]) for record in self.records),
|
|
@@ -446,8 +474,9 @@ class SchemaAndCountConsistencyTest(unittest.TestCase):
|
|
| 446 |
)
|
| 447 |
|
| 448 |
def test_excluded_commercial_fields_did_not_leak(self):
|
| 449 |
-
blob =
|
| 450 |
-
|
|
|
|
| 451 |
for field in (
|
| 452 |
"pricing_entry",
|
| 453 |
"choose_if",
|
|
@@ -461,6 +490,192 @@ class SchemaAndCountConsistencyTest(unittest.TestCase):
|
|
| 461 |
self.assertNotIn(field, blob, f"{field} leaked into the published data")
|
| 462 |
|
| 463 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 464 |
class RankingIntegrityTest(unittest.TestCase):
|
| 465 |
def setUp(self):
|
| 466 |
self.records = read_jsonl_records()
|
|
@@ -519,12 +734,9 @@ class UrlPolicyTest(unittest.TestCase):
|
|
| 519 |
URL_FIELDS = ("provider_page_url", "source_url")
|
| 520 |
|
| 521 |
def test_data_urls_are_https(self):
|
| 522 |
-
for row in read_csv_rows():
|
| 523 |
for field in self.URL_FIELDS:
|
| 524 |
self.assertTrue(row[field].startswith("https://"), f"{field}={row[field]}")
|
| 525 |
-
for record in read_jsonl_records():
|
| 526 |
-
for field in self.URL_FIELDS:
|
| 527 |
-
self.assertTrue(record[field].startswith("https://"), f"{field}={record[field]}")
|
| 528 |
|
| 529 |
def test_summary_urls_are_https(self):
|
| 530 |
summary = read_summary()
|
|
@@ -548,8 +760,8 @@ class UrlPolicyTest(unittest.TestCase):
|
|
| 548 |
self.assertEqual(offenders, [])
|
| 549 |
|
| 550 |
def test_data_files_carry_no_campaign_tagged_urls(self):
|
| 551 |
-
for
|
| 552 |
-
self.assertNotIn("utm_",
|
| 553 |
|
| 554 |
|
| 555 |
class ReadmeCardTest(unittest.TestCase):
|
|
@@ -879,6 +1091,11 @@ class PublicUploadSetTest(unittest.TestCase):
|
|
| 879 |
)
|
| 880 |
self.assertIn("LICENSE", makefile, "make hashes must cover LICENSE")
|
| 881 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 882 |
|
| 883 |
class NoSecretsOrStandInsTest(unittest.TestCase):
|
| 884 |
"""Patterns are written self-avoidingly so this file can be scanned by its own rules."""
|
|
@@ -922,7 +1139,9 @@ class NoSecretsOrStandInsTest(unittest.TestCase):
|
|
| 922 |
self.assertEqual(hits, [])
|
| 923 |
|
| 924 |
def test_no_environment_or_key_material_is_referenced_in_the_data(self):
|
| 925 |
-
blob =
|
|
|
|
|
|
|
| 926 |
for needle in ("Bearer", "api.nativeport.ai", "accounts.nativeport.ai", "panel."):
|
| 927 |
self.assertNotIn(needle, blob, needle)
|
| 928 |
|
|
|
|
| 27 |
ROOT = Path(__file__).resolve().parent.parent
|
| 28 |
DATA = ROOT / "data"
|
| 29 |
CSV_PATH = DATA / "benchmarks.csv"
|
| 30 |
+
METRIC_ROWS_PATH = DATA / "metric_rows.jsonl"
|
| 31 |
JSONL_PATH = DATA / "benchmarks.jsonl"
|
| 32 |
SUMMARY_PATH = DATA / "summary.json"
|
| 33 |
+
|
| 34 |
+
DATA_FILE_NAMES = ("metric_rows.jsonl", "benchmarks.csv", "benchmarks.jsonl", "summary.json")
|
| 35 |
+
|
| 36 |
+
# The Hugging Face configs declared in the card's front matter and the file backing
|
| 37 |
+
# each. Both must be JSONL: the Hub resolves one packaged loader for the whole
|
| 38 |
+
# repository from the declared config data files and applies it to every config, so
|
| 39 |
+
# mixing formats parses one config with the other's reader. The CSV backs no config.
|
| 40 |
+
HUB_CONFIG_DATA_FILES = {
|
| 41 |
+
"metric_rows": "data/metric_rows.jsonl",
|
| 42 |
+
"evaluations": "data/benchmarks.jsonl",
|
| 43 |
+
}
|
| 44 |
README_PATH = ROOT / "README.md"
|
| 45 |
LICENSE_PATH = ROOT / "LICENSE"
|
| 46 |
MAKEFILE_PATH = ROOT / "Makefile"
|
|
|
|
| 75 |
PUBLIC_UPLOAD_SET = (
|
| 76 |
"README.md",
|
| 77 |
"LICENSE",
|
| 78 |
+
"data/metric_rows.jsonl",
|
| 79 |
"data/benchmarks.csv",
|
| 80 |
"data/benchmarks.jsonl",
|
| 81 |
"data/summary.json",
|
|
|
|
| 116 |
return [json.loads(line) for line in handle]
|
| 117 |
|
| 118 |
|
| 119 |
+
def read_metric_row_records():
|
| 120 |
+
with METRIC_ROWS_PATH.open(encoding="utf-8") as handle:
|
| 121 |
+
return [json.loads(line) for line in handle]
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def csv_text_of(value):
|
| 125 |
+
"""The CSV rendering of one typed tidy value, per the generator's contract."""
|
| 126 |
+
if isinstance(value, bool):
|
| 127 |
+
return "true" if value else "false"
|
| 128 |
+
if isinstance(value, (int, float)):
|
| 129 |
+
return json.dumps(value)
|
| 130 |
+
return value
|
| 131 |
+
|
| 132 |
+
|
| 133 |
def read_summary():
|
| 134 |
return json.loads(SUMMARY_PATH.read_text(encoding="utf-8"))
|
| 135 |
|
|
|
|
| 257 |
digests.append(
|
| 258 |
tuple(
|
| 259 |
sha256_of(out / name)
|
| 260 |
+
for name in DATA_FILE_NAMES
|
| 261 |
)
|
| 262 |
)
|
| 263 |
self.assertEqual(digests[0], digests[1])
|
|
|
|
| 308 |
with tempfile.TemporaryDirectory() as workspace:
|
| 309 |
out = Path(workspace) / "out"
|
| 310 |
build_dataset.build(SOURCE_PATH, out)
|
| 311 |
+
for name in DATA_FILE_NAMES:
|
| 312 |
self.assertEqual(
|
| 313 |
sha256_of(out / name),
|
| 314 |
sha256_of(DATA / name),
|
|
|
|
| 392 |
def test_csv_header_matches_the_generator_contract(self):
|
| 393 |
with CSV_PATH.open(encoding="utf-8", newline="") as handle:
|
| 394 |
header = next(csv.reader(handle))
|
| 395 |
+
self.assertEqual(header, build_dataset.TIDY_FIELDS)
|
| 396 |
|
| 397 |
def test_row_counts_agree_across_artifacts(self):
|
| 398 |
self.assertEqual(len(self.rows), self.summary["metric_row_count"])
|
| 399 |
+
self.assertEqual(len(read_metric_row_records()), self.summary["metric_row_count"])
|
| 400 |
self.assertEqual(len(self.records), self.summary["evaluation_count"])
|
| 401 |
self.assertEqual(
|
| 402 |
sum(len(record["metrics"]) for record in self.records),
|
|
|
|
| 474 |
)
|
| 475 |
|
| 476 |
def test_excluded_commercial_fields_did_not_leak(self):
|
| 477 |
+
blob = "".join(
|
| 478 |
+
(DATA / name).read_text(encoding="utf-8") for name in DATA_FILE_NAMES
|
| 479 |
+
)
|
| 480 |
for field in (
|
| 481 |
"pricing_entry",
|
| 482 |
"choose_if",
|
|
|
|
| 490 |
self.assertNotIn(field, blob, f"{field} leaked into the published data")
|
| 491 |
|
| 492 |
|
| 493 |
+
class HubConfigFormatTest(unittest.TestCase):
|
| 494 |
+
"""Both declared configs must be backed by the same data format.
|
| 495 |
+
|
| 496 |
+
The Hub resolves one packaged loader for the whole repository from the declared
|
| 497 |
+
config data files and applies it to every config. A repository that mixes CSV and
|
| 498 |
+
JSONL across configs therefore has one config parsed by the other's reader:
|
| 499 |
+
`load_dataset(..., "evaluations")` failed on the published revision, and the
|
| 500 |
+
datasets-server first-rows call for it failed with it. These tests pin the fix.
|
| 501 |
+
"""
|
| 502 |
+
|
| 503 |
+
def setUp(self):
|
| 504 |
+
self.front = README_PATH.read_text(encoding="utf-8")
|
| 505 |
+
self.front = self.front[4 : self.front.find("\n---\n", 4)]
|
| 506 |
+
|
| 507 |
+
def declared_configs(self):
|
| 508 |
+
"""(config_name, path) pairs read from the front matter, in declared order."""
|
| 509 |
+
pairs, config_name = [], None
|
| 510 |
+
for line in self.front.splitlines():
|
| 511 |
+
stripped = line.strip()
|
| 512 |
+
if stripped.startswith("- config_name:"):
|
| 513 |
+
config_name = stripped.split(":", 1)[1].strip()
|
| 514 |
+
elif stripped.startswith("path:"):
|
| 515 |
+
pairs.append((config_name, stripped.split(":", 1)[1].strip()))
|
| 516 |
+
return pairs
|
| 517 |
+
|
| 518 |
+
def test_front_matter_declares_the_expected_config_to_file_mapping(self):
|
| 519 |
+
self.assertEqual(dict(self.declared_configs()), HUB_CONFIG_DATA_FILES)
|
| 520 |
+
self.assertEqual(self.declared_configs(), list(HUB_CONFIG_DATA_FILES.items()))
|
| 521 |
+
|
| 522 |
+
def test_generator_and_card_agree_on_the_config_mapping(self):
|
| 523 |
+
self.assertEqual(build_dataset.HUB_CONFIG_DATA_FILES, HUB_CONFIG_DATA_FILES)
|
| 524 |
+
self.assertEqual(read_summary()["hub_config_data_files"], HUB_CONFIG_DATA_FILES)
|
| 525 |
+
|
| 526 |
+
def test_every_config_data_file_is_jsonl(self):
|
| 527 |
+
"""The regression itself: one format across all declared configs."""
|
| 528 |
+
suffixes = {Path(path).suffix for _, path in self.declared_configs()}
|
| 529 |
+
self.assertEqual(suffixes, {".jsonl"}, f"configs mix formats: {suffixes}")
|
| 530 |
+
for _, relative in self.declared_configs():
|
| 531 |
+
self.assertTrue((ROOT / relative).is_file(), relative)
|
| 532 |
+
|
| 533 |
+
def test_no_config_points_at_the_csv(self):
|
| 534 |
+
for _, relative in self.declared_configs():
|
| 535 |
+
self.assertNotEqual(relative, "data/benchmarks.csv")
|
| 536 |
+
self.assertEqual(
|
| 537 |
+
read_summary()["downloadable_only_artifacts"], ["data/benchmarks.csv"]
|
| 538 |
+
)
|
| 539 |
+
|
| 540 |
+
def test_every_config_data_file_parses_as_json_lines(self):
|
| 541 |
+
"""Not just the extension: every line of every config file is one JSON object."""
|
| 542 |
+
for config_name, relative in self.declared_configs():
|
| 543 |
+
path = ROOT / relative
|
| 544 |
+
with path.open(encoding="utf-8") as handle:
|
| 545 |
+
lines = handle.read().splitlines()
|
| 546 |
+
self.assertTrue(lines, relative)
|
| 547 |
+
for number, line in enumerate(lines, start=1):
|
| 548 |
+
try:
|
| 549 |
+
record = json.loads(line)
|
| 550 |
+
except json.JSONDecodeError as exc:
|
| 551 |
+
self.fail(f"{config_name} ({relative}) line {number}: {exc}")
|
| 552 |
+
self.assertIsInstance(record, dict, f"{relative} line {number}")
|
| 553 |
+
|
| 554 |
+
def test_config_records_are_uniform_in_fields_and_types(self):
|
| 555 |
+
"""One type per field across every record.
|
| 556 |
+
|
| 557 |
+
The loader infers one Arrow column type per field from the records it reads. A
|
| 558 |
+
field that is an integer on one line and a float or a string on another is the
|
| 559 |
+
other common way a config loads locally but fails the viewer's first-rows call.
|
| 560 |
+
"""
|
| 561 |
+
for config_name, relative in self.declared_configs():
|
| 562 |
+
lines = (ROOT / relative).read_text(encoding="utf-8").splitlines()
|
| 563 |
+
records = [json.loads(line) for line in lines]
|
| 564 |
+
types = {}
|
| 565 |
+
for index, record in enumerate(records):
|
| 566 |
+
self.assertEqual(
|
| 567 |
+
list(record),
|
| 568 |
+
list(records[0]),
|
| 569 |
+
f"{config_name} record {index}: field set or order differs",
|
| 570 |
+
)
|
| 571 |
+
for field, value in record.items():
|
| 572 |
+
types.setdefault(field, set()).add(type(value).__name__)
|
| 573 |
+
mixed = {field: sorted(seen) for field, seen in types.items() if len(seen) > 1}
|
| 574 |
+
self.assertEqual(mixed, {}, f"{config_name}: fields change type between records")
|
| 575 |
+
|
| 576 |
+
def test_csv_is_still_published_as_a_downloadable_artifact(self):
|
| 577 |
+
self.assertTrue(CSV_PATH.is_file())
|
| 578 |
+
self.assertIn("benchmarks.csv", read_summary()["outputs"])
|
| 579 |
+
self.assertIn("data/benchmarks.csv", PUBLIC_UPLOAD_SET)
|
| 580 |
+
readme = " ".join(README_PATH.read_text(encoding="utf-8").split())
|
| 581 |
+
self.assertIn("`data/benchmarks.csv` is a download rather than a config", readme)
|
| 582 |
+
self.assertIn("| `data/benchmarks.csv` | 297 |", readme)
|
| 583 |
+
|
| 584 |
+
|
| 585 |
+
class TidyMirrorEquivalenceTest(unittest.TestCase):
|
| 586 |
+
"""data/benchmarks.csv and data/metric_rows.jsonl are the same tidy rows.
|
| 587 |
+
|
| 588 |
+
The CSV is the approved artifact; the JSONL is what the `metric_rows` config loads.
|
| 589 |
+
They are written from one row builder, and these tests hold them to that: same
|
| 590 |
+
fields in the same order, same count, same order of rows, and every value equal —
|
| 591 |
+
numbers compared as text, so a rounded or reformatted measurement fails here.
|
| 592 |
+
"""
|
| 593 |
+
|
| 594 |
+
def setUp(self):
|
| 595 |
+
self.rows = read_csv_rows()
|
| 596 |
+
self.records = read_metric_row_records()
|
| 597 |
+
|
| 598 |
+
def test_both_tidy_views_carry_the_same_row_count(self):
|
| 599 |
+
summary = read_summary()
|
| 600 |
+
self.assertEqual(len(self.rows), summary["metric_row_count"])
|
| 601 |
+
self.assertEqual(len(self.records), summary["metric_row_count"])
|
| 602 |
+
self.assertEqual(len(self.rows), len(self.records))
|
| 603 |
+
|
| 604 |
+
def test_field_names_and_order_are_identical(self):
|
| 605 |
+
with CSV_PATH.open(encoding="utf-8", newline="") as handle:
|
| 606 |
+
header = next(csv.reader(handle))
|
| 607 |
+
self.assertEqual(header, build_dataset.TIDY_FIELDS)
|
| 608 |
+
for index, record in enumerate(self.records):
|
| 609 |
+
self.assertEqual(list(record), build_dataset.TIDY_FIELDS, f"record {index}")
|
| 610 |
+
|
| 611 |
+
def test_rows_are_semantically_identical_in_the_same_order(self):
|
| 612 |
+
for index, (row, record) in enumerate(zip(self.rows, self.records)):
|
| 613 |
+
self.assertEqual(
|
| 614 |
+
(record["evaluation_id"], record["metric_key"]),
|
| 615 |
+
(row["evaluation_id"], row["metric_key"]),
|
| 616 |
+
f"row {index}: the two tidy views are ordered differently",
|
| 617 |
+
)
|
| 618 |
+
for field in build_dataset.TIDY_FIELDS:
|
| 619 |
+
self.assertEqual(
|
| 620 |
+
csv_text_of(record[field]),
|
| 621 |
+
row[field],
|
| 622 |
+
f"row {index} field {field}",
|
| 623 |
+
)
|
| 624 |
+
|
| 625 |
+
def test_numbers_are_preserved_exactly_not_reformatted(self):
|
| 626 |
+
"""Every numeric field is a JSON number whose text is the CSV token verbatim."""
|
| 627 |
+
numeric = ("metric_raw", "composite_score", "metric_index", "rank", "rank_of",
|
| 628 |
+
"composite_scale_max")
|
| 629 |
+
compared = 0
|
| 630 |
+
for row, record in zip(self.rows, self.records):
|
| 631 |
+
for field in numeric:
|
| 632 |
+
value = record[field]
|
| 633 |
+
self.assertIsInstance(value, (int, float), field)
|
| 634 |
+
self.assertNotIsInstance(value, bool, field)
|
| 635 |
+
self.assertEqual(json.dumps(value), row[field], f"{field} in {row['evaluation_id']}")
|
| 636 |
+
compared += 1
|
| 637 |
+
self.assertIsInstance(record["is_capability_top"], bool)
|
| 638 |
+
self.assertEqual(csv_text_of(record["is_capability_top"]), row["is_capability_top"])
|
| 639 |
+
self.assertEqual(compared, len(self.rows) * len(numeric))
|
| 640 |
+
|
| 641 |
+
def test_tidy_jsonl_agrees_with_the_nested_evaluations(self):
|
| 642 |
+
"""The third view must carry the same measurements as the other two."""
|
| 643 |
+
nested = {
|
| 644 |
+
record["evaluation_id"]: {m["metric_key"]: m for m in record["metrics"]}
|
| 645 |
+
for record in read_jsonl_records()
|
| 646 |
+
}
|
| 647 |
+
seen = set()
|
| 648 |
+
for record in self.records:
|
| 649 |
+
metric = nested[record["evaluation_id"]][record["metric_key"]]
|
| 650 |
+
self.assertEqual(json.dumps(record["metric_raw"]), json.dumps(metric["raw_value"]))
|
| 651 |
+
self.assertEqual(record["metric_display"], metric["display_value"])
|
| 652 |
+
self.assertEqual(record["metric_label"], metric["metric_label"])
|
| 653 |
+
self.assertEqual(record["metric_index"], metric["metric_index"])
|
| 654 |
+
seen.add((record["evaluation_id"], record["metric_key"]))
|
| 655 |
+
self.assertEqual(len(seen), len(self.records))
|
| 656 |
+
self.assertEqual(
|
| 657 |
+
len(seen), sum(len(metrics) for metrics in nested.values())
|
| 658 |
+
)
|
| 659 |
+
|
| 660 |
+
def test_generator_emits_both_tidy_views_from_one_builder(self):
|
| 661 |
+
"""A fixture build: change nothing, and the two files still agree row for row."""
|
| 662 |
+
with tempfile.TemporaryDirectory() as workspace:
|
| 663 |
+
out = Path(workspace) / "out"
|
| 664 |
+
build_dataset.build(write_fixture(workspace), out)
|
| 665 |
+
with (out / "benchmarks.csv").open(encoding="utf-8", newline="") as handle:
|
| 666 |
+
rows = list(csv.DictReader(handle))
|
| 667 |
+
records = [
|
| 668 |
+
json.loads(line)
|
| 669 |
+
for line in (out / "metric_rows.jsonl").read_text(encoding="utf-8").splitlines()
|
| 670 |
+
]
|
| 671 |
+
self.assertEqual(len(rows), len(records))
|
| 672 |
+
self.assertEqual(len(rows), 10) # 3 + 4 + 3 metrics in the fixture
|
| 673 |
+
for row, record in zip(rows, records):
|
| 674 |
+
self.assertEqual(list(record), build_dataset.TIDY_FIELDS)
|
| 675 |
+
for field in build_dataset.TIDY_FIELDS:
|
| 676 |
+
self.assertEqual(csv_text_of(record[field]), row[field], field)
|
| 677 |
+
|
| 678 |
+
|
| 679 |
class RankingIntegrityTest(unittest.TestCase):
|
| 680 |
def setUp(self):
|
| 681 |
self.records = read_jsonl_records()
|
|
|
|
| 734 |
URL_FIELDS = ("provider_page_url", "source_url")
|
| 735 |
|
| 736 |
def test_data_urls_are_https(self):
|
| 737 |
+
for row in read_csv_rows() + read_metric_row_records() + read_jsonl_records():
|
| 738 |
for field in self.URL_FIELDS:
|
| 739 |
self.assertTrue(row[field].startswith("https://"), f"{field}={row[field]}")
|
|
|
|
|
|
|
|
|
|
| 740 |
|
| 741 |
def test_summary_urls_are_https(self):
|
| 742 |
summary = read_summary()
|
|
|
|
| 760 |
self.assertEqual(offenders, [])
|
| 761 |
|
| 762 |
def test_data_files_carry_no_campaign_tagged_urls(self):
|
| 763 |
+
for name in DATA_FILE_NAMES:
|
| 764 |
+
self.assertNotIn("utm_", (DATA / name).read_text(encoding="utf-8"), name)
|
| 765 |
|
| 766 |
|
| 767 |
class ReadmeCardTest(unittest.TestCase):
|
|
|
|
| 1091 |
)
|
| 1092 |
self.assertIn("LICENSE", makefile, "make hashes must cover LICENSE")
|
| 1093 |
|
| 1094 |
+
def test_make_hashes_covers_every_data_file(self):
|
| 1095 |
+
makefile = MAKEFILE_PATH.read_text(encoding="utf-8")
|
| 1096 |
+
for name in DATA_FILE_NAMES:
|
| 1097 |
+
self.assertIn(name, makefile, f"make hashes must cover {name}")
|
| 1098 |
+
|
| 1099 |
|
| 1100 |
class NoSecretsOrStandInsTest(unittest.TestCase):
|
| 1101 |
"""Patterns are written self-avoidingly so this file can be scanned by its own rules."""
|
|
|
|
| 1139 |
self.assertEqual(hits, [])
|
| 1140 |
|
| 1141 |
def test_no_environment_or_key_material_is_referenced_in_the_data(self):
|
| 1142 |
+
blob = "".join(
|
| 1143 |
+
(DATA / name).read_text(encoding="utf-8") for name in DATA_FILE_NAMES
|
| 1144 |
+
)
|
| 1145 |
for needle in ("Bearer", "api.nativeport.ai", "accounts.nativeport.ai", "panel."):
|
| 1146 |
self.assertNotIn(needle, blob, needle)
|
| 1147 |
|