sable-sphynx 0.7.0.dev0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- sable_sphynx-0.7.0.dev0.dist-info/METADATA +152 -0
- sable_sphynx-0.7.0.dev0.dist-info/RECORD +174 -0
- sable_sphynx-0.7.0.dev0.dist-info/WHEEL +5 -0
- sable_sphynx-0.7.0.dev0.dist-info/top_level.txt +1 -0
- sphynx/__init__.py +469 -0
- sphynx/__init__.pyi +376 -0
- sphynx/_analytical_client.py +77 -0
- sphynx/_analytical_values.py +252 -0
- sphynx/_audit_client.py +173 -0
- sphynx/_comparison_adapters.py +113 -0
- sphynx/_componentized_full_rag.py +34 -0
- sphynx/_composed_pipeline.py +109 -0
- sphynx/_full_rag_grading.py +218 -0
- sphynx/_full_rag_trace.py +221 -0
- sphynx/_local_eval.py +268 -0
- sphynx/_pipeline_core.py +243 -0
- sphynx/_pr_client.py +237 -0
- sphynx/_rag_diagnostics_calibration.py +285 -0
- sphynx/_rag_diagnostics_catalog.py +39 -0
- sphynx/_rag_diagnostics_common.py +26 -0
- sphynx/_rag_diagnostics_contracts.py +186 -0
- sphynx/_rag_diagnostics_engine.py +243 -0
- sphynx/_rag_diagnostics_importers.py +242 -0
- sphynx/_rag_diagnostics_ranking.py +81 -0
- sphynx/_rag_diagnostics_semantic.py +325 -0
- sphynx/_rag_ranking.py +149 -0
- sphynx/_speech_benchmark_scoring.py +140 -0
- sphynx/_speech_benchmark_sources.py +275 -0
- sphynx/_speech_complete_scoring.py +218 -0
- sphynx/_speech_complete_source.py +118 -0
- sphynx/_speech_kido.py +294 -0
- sphynx/_speech_scoring.py +77 -0
- sphynx/_track_capture.py +213 -0
- sphynx/analytical_views.py +515 -0
- sphynx/assessment_contracts.py +378 -0
- sphynx/assessment_selection.py +57 -0
- sphynx/audit_access.py +245 -0
- sphynx/audit_contracts.py +204 -0
- sphynx/audit_requests.py +180 -0
- sphynx/authoring.py +188 -0
- sphynx/benchmarks.py +568 -0
- sphynx/candidates.py +62 -0
- sphynx/catalog.py +272 -0
- sphynx/chunking.py +372 -0
- sphynx/clickhouse_reader.py +169 -0
- sphynx/client.py +486 -0
- sphynx/comparison_core.py +149 -0
- sphynx/comparison_views.py +223 -0
- sphynx/components.py +1661 -0
- sphynx/contracts.py +514 -0
- sphynx/customer_rerank.py +299 -0
- sphynx/customer_retrieval.py +146 -0
- sphynx/data/benchmark-releases.v1.json +164 -0
- sphynx/data/public-benchmark-release.v1.json +354 -0
- sphynx/data/rag-diagnostics-synthetic.v1.json +6982 -0
- sphynx/data/rag-paper-references.v1.json +176 -0
- sphynx/data/rag-paper-schemas.v1.json +1069 -0
- sphynx/data/rag-rerank-v1.json +507 -0
- sphynx/dataset_catalog_data.py +1214 -0
- sphynx/dataset_readiness.py +138 -0
- sphynx/dataset_v3.py +999 -0
- sphynx/demo.py +130 -0
- sphynx/demo_candidates/__init__.py +1 -0
- sphynx/demo_candidates/dense.py +11 -0
- sphynx/demo_candidates/lexical.py +21 -0
- sphynx/diagnostics.py +207 -0
- sphynx/documentation.py +737 -0
- sphynx/documentation_candidate.py +939 -0
- sphynx/evals.py +538 -0
- sphynx/evaluation_protocols.py +124 -0
- sphynx/evaluator_bindings.py +370 -0
- sphynx/evidence.py +352 -0
- sphynx/execution_capabilities.py +109 -0
- sphynx/fast_search.py +285 -0
- sphynx/full_rag.py +341 -0
- sphynx/hosted/__init__.py +18 -0
- sphynx/hosted/benchmark_intake.py +218 -0
- sphynx/hosted/contracts.py +361 -0
- sphynx/hosted/model_admission.py +130 -0
- sphynx/hosted/model_artifacts.py +298 -0
- sphynx/hosted/model_contracts.py +63 -0
- sphynx/hosted/model_gcs.py +181 -0
- sphynx/hosted/model_history.py +84 -0
- sphynx/hosted/model_session.py +233 -0
- sphynx/hosted/public_suite.py +74 -0
- sphynx/hosted/registry.py +97 -0
- sphynx/hosted/revisions.py +126 -0
- sphynx/hosted/sandbox.py +105 -0
- sphynx/hosted/submission.py +287 -0
- sphynx/hosted_workflow.py +451 -0
- sphynx/http_transport.py +86 -0
- sphynx/integrated_rag.py +352 -0
- sphynx/integrated_workflow.py +552 -0
- sphynx/integration_contracts.py +474 -0
- sphynx/invocation.py +791 -0
- sphynx/invocation_candidates.py +392 -0
- sphynx/invocation_catalog_data.py +40 -0
- sphynx/learned_catalog_data.py +30 -0
- sphynx/learned_rerank.py +267 -0
- sphynx/legacy_full_rag.py +67 -0
- sphynx/lifecycle.py +885 -0
- sphynx/lifecycle_workflow.py +490 -0
- sphynx/metric_views.py +474 -0
- sphynx/metrics.py +282 -0
- sphynx/model_adapters.py +438 -0
- sphynx/model_catalog_data.py +30 -0
- sphynx/official_benchmarks.py +385 -0
- sphynx/packs.py +258 -0
- sphynx/pipeline.py +313 -0
- sphynx/pipeline_workflow.py +551 -0
- sphynx/pr_evaluations.py +560 -0
- sphynx/progress.py +89 -0
- sphynx/public_bundle.py +173 -0
- sphynx/public_eval.py +62 -0
- sphynx/public_rerank.py +355 -0
- sphynx/py.typed +0 -0
- sphynx/pytest_plugin.py +62 -0
- sphynx/question_robustness.py +437 -0
- sphynx/rag.py +329 -0
- sphynx/rag_diagnostics.py +309 -0
- sphynx/rag_papers/__init__.py +29 -0
- sphynx/rag_papers/__main__.py +45 -0
- sphynx/rag_papers/candidates.py +19 -0
- sphynx/rag_papers/contracts.py +282 -0
- sphynx/rag_papers/datasets.py +291 -0
- sphynx/rag_papers/experiments.py +116 -0
- sphynx/rag_papers/judges.py +308 -0
- sphynx/rag_papers/registry.py +33 -0
- sphynx/rag_papers/runner.py +176 -0
- sphynx/rag_papers/sable.py +135 -0
- sphynx/rag_papers/schemas.py +12 -0
- sphynx/rag_papers/scoring.py +249 -0
- sphynx/rag_papers/semantic.py +288 -0
- sphynx/rag_papers/trec.py +80 -0
- sphynx/recorded_scores.py +352 -0
- sphynx/records.py +124 -0
- sphynx/reflex_traces.py +193 -0
- sphynx/report_rendering.py +71 -0
- sphynx/rerank_family_contracts.py +150 -0
- sphynx/rerank_workflow.py +424 -0
- sphynx/research_candidate.py +190 -0
- sphynx/results.py +852 -0
- sphynx/retrieval.py +750 -0
- sphynx/retrieval_candidates.py +295 -0
- sphynx/retrieval_catalog_data.py +19 -0
- sphynx/retrieval_recordings.py +125 -0
- sphynx/reviewer_audit.py +266 -0
- sphynx/routing_policies.py +375 -0
- sphynx/runner.py +544 -0
- sphynx/sable_rerank.py +489 -0
- sphynx/score_reliability.py +393 -0
- sphynx/speech_benchmarks.py +204 -0
- sphynx/speech_campaigns.py +378 -0
- sphynx/speech_complete.py +236 -0
- sphynx/speech_contracts.py +225 -0
- sphynx/speech_evidence.py +429 -0
- sphynx/speech_listening.py +115 -0
- sphynx/speech_multilingual.py +71 -0
- sphynx/speech_tts.py +479 -0
- sphynx/sql/001_sync_v1.sql +30 -0
- sphynx/stage_contracts.py +199 -0
- sphynx/statistical_catalog_data.py +53 -0
- sphynx/statistical_rerank.py +232 -0
- sphynx/sync.py +797 -0
- sphynx/workbench.py +378 -0
- sphynx/workbench_catalog_data.py +13 -0
- sphynx/workbench_tracks.py +78 -0
- sphynx/workflow_async.py +197 -0
- sphynx/workflow_contracts.py +594 -0
- sphynx/workflow_decisions.py +211 -0
- sphynx/workflow_defaults.py +16 -0
- sphynx/workflow_extensions.py +212 -0
- sphynx/workflow_registry.py +513 -0
- sphynx/workflows.py +792 -0
|
@@ -0,0 +1,354 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schema_version": "sphynx.hosted.public-admission.v1",
|
|
3
|
+
"manifest_digest": "sha256:9335ce0aaa3e53942c8cba2deff126af5ad93319d383cb4eb70fceb443990835",
|
|
4
|
+
"benchmark_release": {
|
|
5
|
+
"cases": [
|
|
6
|
+
{
|
|
7
|
+
"case_id": "scifact:577",
|
|
8
|
+
"company": "scifact",
|
|
9
|
+
"form_receipt": null,
|
|
10
|
+
"form_status": "unclassified",
|
|
11
|
+
"input_digest": "sha256:ca86c3a18b08ba81a462467fde1d35d19fff96fef7f0fa7bbb8a5211342f4259",
|
|
12
|
+
"input_form": "unclassified",
|
|
13
|
+
"length_words": 24,
|
|
14
|
+
"split": "development"
|
|
15
|
+
},
|
|
16
|
+
{
|
|
17
|
+
"case_id": "scifact:528",
|
|
18
|
+
"company": "scifact",
|
|
19
|
+
"form_receipt": null,
|
|
20
|
+
"form_status": "unclassified",
|
|
21
|
+
"input_digest": "sha256:6f03bc521d6c5a75d3b27f8b208671deb20c6c2590c34a613b3ce80caad35436",
|
|
22
|
+
"input_form": "unclassified",
|
|
23
|
+
"length_words": 24,
|
|
24
|
+
"split": "development"
|
|
25
|
+
},
|
|
26
|
+
{
|
|
27
|
+
"case_id": "scifact:1089",
|
|
28
|
+
"company": "scifact",
|
|
29
|
+
"form_receipt": null,
|
|
30
|
+
"form_status": "unclassified",
|
|
31
|
+
"input_digest": "sha256:133dfc71d31d3a0e0eb27810e8d19f7837850f25cd82b70b1392dd87c7633fff",
|
|
32
|
+
"input_form": "unclassified",
|
|
33
|
+
"length_words": 13,
|
|
34
|
+
"split": "development"
|
|
35
|
+
},
|
|
36
|
+
{
|
|
37
|
+
"case_id": "scifact:298",
|
|
38
|
+
"company": "scifact",
|
|
39
|
+
"form_receipt": null,
|
|
40
|
+
"form_status": "unclassified",
|
|
41
|
+
"input_digest": "sha256:94e5a151f645c3e69218af7cb070adc2a0bb7ae2abb39a43c7f249618f9c305a",
|
|
42
|
+
"input_form": "unclassified",
|
|
43
|
+
"length_words": 13,
|
|
44
|
+
"split": "development"
|
|
45
|
+
},
|
|
46
|
+
{
|
|
47
|
+
"case_id": "scifact:578",
|
|
48
|
+
"company": "scifact",
|
|
49
|
+
"form_receipt": null,
|
|
50
|
+
"form_status": "unclassified",
|
|
51
|
+
"input_digest": "sha256:f7a0ac8e2038bf3814cb7461ab1b1c1bad4c7bb99a3de62df8d8cfe2d56cb7bb",
|
|
52
|
+
"input_form": "unclassified",
|
|
53
|
+
"length_words": 10,
|
|
54
|
+
"split": "development"
|
|
55
|
+
},
|
|
56
|
+
{
|
|
57
|
+
"case_id": "scifact:721",
|
|
58
|
+
"company": "scifact",
|
|
59
|
+
"form_receipt": null,
|
|
60
|
+
"form_status": "unclassified",
|
|
61
|
+
"input_digest": "sha256:8b0da963df01fee6cdb9ba1a34c390857675ca45497a0b6c00546e85f1d22f5d",
|
|
62
|
+
"input_form": "unclassified",
|
|
63
|
+
"length_words": 13,
|
|
64
|
+
"split": "development"
|
|
65
|
+
},
|
|
66
|
+
{
|
|
67
|
+
"case_id": "scifact:1180",
|
|
68
|
+
"company": "scifact",
|
|
69
|
+
"form_receipt": null,
|
|
70
|
+
"form_status": "unclassified",
|
|
71
|
+
"input_digest": "sha256:0283cc6b0ccf76c9616b396e5a29d8e937d2567e6270654e46398e6249f7cf1c",
|
|
72
|
+
"input_form": "unclassified",
|
|
73
|
+
"length_words": 10,
|
|
74
|
+
"split": "development"
|
|
75
|
+
},
|
|
76
|
+
{
|
|
77
|
+
"case_id": "scifact:386",
|
|
78
|
+
"company": "scifact",
|
|
79
|
+
"form_receipt": null,
|
|
80
|
+
"form_status": "unclassified",
|
|
81
|
+
"input_digest": "sha256:9ae61bbf3af9c941f837bb3769939aefa1eebad95961c87ee42c6d0817abb97e",
|
|
82
|
+
"input_form": "unclassified",
|
|
83
|
+
"length_words": 16,
|
|
84
|
+
"split": "development"
|
|
85
|
+
},
|
|
86
|
+
{
|
|
87
|
+
"case_id": "scifact:410",
|
|
88
|
+
"company": "scifact",
|
|
89
|
+
"form_receipt": null,
|
|
90
|
+
"form_status": "unclassified",
|
|
91
|
+
"input_digest": "sha256:f58cd00ff9451b4bad972862dbcdf9439e755a67fdc83dfad3eeca3fc02530de",
|
|
92
|
+
"input_form": "unclassified",
|
|
93
|
+
"length_words": 9,
|
|
94
|
+
"split": "development"
|
|
95
|
+
},
|
|
96
|
+
{
|
|
97
|
+
"case_id": "scifact:1100",
|
|
98
|
+
"company": "scifact",
|
|
99
|
+
"form_receipt": null,
|
|
100
|
+
"form_status": "unclassified",
|
|
101
|
+
"input_digest": "sha256:1df6cafdeb16cd5b5467e9e9c535968ea0f789424d7bdafaceece75d07cfd80a",
|
|
102
|
+
"input_form": "unclassified",
|
|
103
|
+
"length_words": 4,
|
|
104
|
+
"split": "development"
|
|
105
|
+
},
|
|
106
|
+
{
|
|
107
|
+
"case_id": "scifact:729",
|
|
108
|
+
"company": "scifact",
|
|
109
|
+
"form_receipt": null,
|
|
110
|
+
"form_status": "unclassified",
|
|
111
|
+
"input_digest": "sha256:670d43c0f5769f1bb4069ec48050e3eb69f58fc42b86f2987440984ceb592b0d",
|
|
112
|
+
"input_form": "unclassified",
|
|
113
|
+
"length_words": 11,
|
|
114
|
+
"split": "development"
|
|
115
|
+
},
|
|
116
|
+
{
|
|
117
|
+
"case_id": "scifact:1335",
|
|
118
|
+
"company": "scifact",
|
|
119
|
+
"form_receipt": null,
|
|
120
|
+
"form_status": "unclassified",
|
|
121
|
+
"input_digest": "sha256:cd4a9e4b3908f52827e41ba62acfa7e5987557a25ab05f295cbf3ce14f2e53de",
|
|
122
|
+
"input_form": "unclassified",
|
|
123
|
+
"length_words": 9,
|
|
124
|
+
"split": "development"
|
|
125
|
+
},
|
|
126
|
+
{
|
|
127
|
+
"case_id": "scifact:525",
|
|
128
|
+
"company": "scifact",
|
|
129
|
+
"form_receipt": null,
|
|
130
|
+
"form_status": "unclassified",
|
|
131
|
+
"input_digest": "sha256:8b950443e3bb0ad378c6c3b49bf52a201e6788d6c01b691afc201fe1b2f40fc5",
|
|
132
|
+
"input_form": "unclassified",
|
|
133
|
+
"length_words": 20,
|
|
134
|
+
"split": "development"
|
|
135
|
+
},
|
|
136
|
+
{
|
|
137
|
+
"case_id": "scifact:659",
|
|
138
|
+
"company": "scifact",
|
|
139
|
+
"form_receipt": null,
|
|
140
|
+
"form_status": "unclassified",
|
|
141
|
+
"input_digest": "sha256:a387ee970f24fad31cf8a35e61ba535c530230ef2530d5888cf44b088a9e6a6c",
|
|
142
|
+
"input_form": "unclassified",
|
|
143
|
+
"length_words": 7,
|
|
144
|
+
"split": "development"
|
|
145
|
+
},
|
|
146
|
+
{
|
|
147
|
+
"case_id": "scifact:549",
|
|
148
|
+
"company": "scifact",
|
|
149
|
+
"form_receipt": null,
|
|
150
|
+
"form_status": "unclassified",
|
|
151
|
+
"input_digest": "sha256:5b8e3479ffd0a289c2d59973c69ac5082dfc7e80775d34c9d4423aaf21d7ccaa",
|
|
152
|
+
"input_form": "unclassified",
|
|
153
|
+
"length_words": 7,
|
|
154
|
+
"split": "development"
|
|
155
|
+
}
|
|
156
|
+
],
|
|
157
|
+
"corpus": {
|
|
158
|
+
"digest": "sha256:0048c9b73791c6c2a66b58a62d612d993d7699855294878b59606d5ca369ec09",
|
|
159
|
+
"ref": "beir.scifact.pool-corpus.dev15@1.0.0",
|
|
160
|
+
"version": "1.0.0"
|
|
161
|
+
},
|
|
162
|
+
"evaluation_id": "rag.rerank.public.v1",
|
|
163
|
+
"input": {
|
|
164
|
+
"digest": "sha256:3a80c8b09cb26c96c7551d6b194d8195fc848ea7a0491c8f014a30e505bf5140",
|
|
165
|
+
"ref": "beir.scifact.inputs.dev15@1.0.0",
|
|
166
|
+
"version": "1.0.0"
|
|
167
|
+
},
|
|
168
|
+
"intended_use": "development",
|
|
169
|
+
"interface_version": "sphynx.rag.rerank.v1",
|
|
170
|
+
"label_set": {
|
|
171
|
+
"digest": "sha256:18b60b759c05ac9a8d5ac14477ff509e67ec9c38ffd760c20b0b0f54313ad198",
|
|
172
|
+
"ref": "beir.scifact.native-qrels.dev15@1.0.0",
|
|
173
|
+
"version": "1.0.0"
|
|
174
|
+
},
|
|
175
|
+
"metric_profile": "rag.rerank.beir.metric.v1",
|
|
176
|
+
"protocol_revision": "rag.rerank.scifact.dev15.grader.v1",
|
|
177
|
+
"ref": "beir.scifact.bm25-top15.dev15@1.0.0",
|
|
178
|
+
"required_input_form": "any",
|
|
179
|
+
"required_label_profile": "passage_qrels",
|
|
180
|
+
"schema_version": "sphynx.benchmark-release.v1",
|
|
181
|
+
"scope": "component",
|
|
182
|
+
"version": "1.0.0"
|
|
183
|
+
},
|
|
184
|
+
"label_release": {
|
|
185
|
+
"authority": "native_public_qrels",
|
|
186
|
+
"cases": [
|
|
187
|
+
{
|
|
188
|
+
"case_id": "scifact:577",
|
|
189
|
+
"input_digest": "sha256:ca86c3a18b08ba81a462467fde1d35d19fff96fef7f0fa7bbb8a5211342f4259",
|
|
190
|
+
"status": "judged"
|
|
191
|
+
},
|
|
192
|
+
{
|
|
193
|
+
"case_id": "scifact:528",
|
|
194
|
+
"input_digest": "sha256:6f03bc521d6c5a75d3b27f8b208671deb20c6c2590c34a613b3ce80caad35436",
|
|
195
|
+
"status": "judged"
|
|
196
|
+
},
|
|
197
|
+
{
|
|
198
|
+
"case_id": "scifact:1089",
|
|
199
|
+
"input_digest": "sha256:133dfc71d31d3a0e0eb27810e8d19f7837850f25cd82b70b1392dd87c7633fff",
|
|
200
|
+
"status": "judged"
|
|
201
|
+
},
|
|
202
|
+
{
|
|
203
|
+
"case_id": "scifact:298",
|
|
204
|
+
"input_digest": "sha256:94e5a151f645c3e69218af7cb070adc2a0bb7ae2abb39a43c7f249618f9c305a",
|
|
205
|
+
"status": "judged"
|
|
206
|
+
},
|
|
207
|
+
{
|
|
208
|
+
"case_id": "scifact:578",
|
|
209
|
+
"input_digest": "sha256:f7a0ac8e2038bf3814cb7461ab1b1c1bad4c7bb99a3de62df8d8cfe2d56cb7bb",
|
|
210
|
+
"status": "judged"
|
|
211
|
+
},
|
|
212
|
+
{
|
|
213
|
+
"case_id": "scifact:721",
|
|
214
|
+
"input_digest": "sha256:8b0da963df01fee6cdb9ba1a34c390857675ca45497a0b6c00546e85f1d22f5d",
|
|
215
|
+
"status": "judged"
|
|
216
|
+
},
|
|
217
|
+
{
|
|
218
|
+
"case_id": "scifact:1180",
|
|
219
|
+
"input_digest": "sha256:0283cc6b0ccf76c9616b396e5a29d8e937d2567e6270654e46398e6249f7cf1c",
|
|
220
|
+
"status": "judged"
|
|
221
|
+
},
|
|
222
|
+
{
|
|
223
|
+
"case_id": "scifact:386",
|
|
224
|
+
"input_digest": "sha256:9ae61bbf3af9c941f837bb3769939aefa1eebad95961c87ee42c6d0817abb97e",
|
|
225
|
+
"status": "judged"
|
|
226
|
+
},
|
|
227
|
+
{
|
|
228
|
+
"case_id": "scifact:410",
|
|
229
|
+
"input_digest": "sha256:f58cd00ff9451b4bad972862dbcdf9439e755a67fdc83dfad3eeca3fc02530de",
|
|
230
|
+
"status": "judged"
|
|
231
|
+
},
|
|
232
|
+
{
|
|
233
|
+
"case_id": "scifact:1100",
|
|
234
|
+
"input_digest": "sha256:1df6cafdeb16cd5b5467e9e9c535968ea0f789424d7bdafaceece75d07cfd80a",
|
|
235
|
+
"status": "judged"
|
|
236
|
+
},
|
|
237
|
+
{
|
|
238
|
+
"case_id": "scifact:729",
|
|
239
|
+
"input_digest": "sha256:670d43c0f5769f1bb4069ec48050e3eb69f58fc42b86f2987440984ceb592b0d",
|
|
240
|
+
"status": "judged"
|
|
241
|
+
},
|
|
242
|
+
{
|
|
243
|
+
"case_id": "scifact:1335",
|
|
244
|
+
"input_digest": "sha256:cd4a9e4b3908f52827e41ba62acfa7e5987557a25ab05f295cbf3ce14f2e53de",
|
|
245
|
+
"status": "judged"
|
|
246
|
+
},
|
|
247
|
+
{
|
|
248
|
+
"case_id": "scifact:525",
|
|
249
|
+
"input_digest": "sha256:8b950443e3bb0ad378c6c3b49bf52a201e6788d6c01b691afc201fe1b2f40fc5",
|
|
250
|
+
"status": "judged"
|
|
251
|
+
},
|
|
252
|
+
{
|
|
253
|
+
"case_id": "scifact:659",
|
|
254
|
+
"input_digest": "sha256:a387ee970f24fad31cf8a35e61ba535c530230ef2530d5888cf44b088a9e6a6c",
|
|
255
|
+
"status": "judged"
|
|
256
|
+
},
|
|
257
|
+
{
|
|
258
|
+
"case_id": "scifact:549",
|
|
259
|
+
"input_digest": "sha256:5b8e3479ffd0a289c2d59973c69ac5082dfc7e80775d34c9d4423aaf21d7ccaa",
|
|
260
|
+
"status": "judged"
|
|
261
|
+
}
|
|
262
|
+
],
|
|
263
|
+
"corpus": {
|
|
264
|
+
"digest": "sha256:0048c9b73791c6c2a66b58a62d612d993d7699855294878b59606d5ca369ec09",
|
|
265
|
+
"ref": "beir.scifact.pool-corpus.dev15@1.0.0",
|
|
266
|
+
"version": "1.0.0"
|
|
267
|
+
},
|
|
268
|
+
"input": {
|
|
269
|
+
"digest": "sha256:3a80c8b09cb26c96c7551d6b194d8195fc848ea7a0491c8f014a30e505bf5140",
|
|
270
|
+
"ref": "beir.scifact.inputs.dev15@1.0.0",
|
|
271
|
+
"version": "1.0.0"
|
|
272
|
+
},
|
|
273
|
+
"label_profile": "passage_qrels",
|
|
274
|
+
"lineage": [],
|
|
275
|
+
"missingness": "native_qrels_unjudged",
|
|
276
|
+
"policy_revision": "native-sparse-qrels.unjudged-as-zero.pool-conditional.v1",
|
|
277
|
+
"ref": "beir.scifact.native-qrels.dev15@1.0.0",
|
|
278
|
+
"reference": {
|
|
279
|
+
"digest": "sha256:4d324d640133fbddc9b75eced88d900e7aed386ddd717a284d30a4b0a72dcf82",
|
|
280
|
+
"ref": "beir.scifact.references.dev15@1.0.0",
|
|
281
|
+
"version": "1.0.0"
|
|
282
|
+
},
|
|
283
|
+
"review_receipts": [],
|
|
284
|
+
"rubric_revision": "rag.rerank.beir.metric.v1",
|
|
285
|
+
"schema_version": "sphynx.label-set-release.v1",
|
|
286
|
+
"version": "1.0.0"
|
|
287
|
+
},
|
|
288
|
+
"population": {
|
|
289
|
+
"case_count": 15,
|
|
290
|
+
"native_judged_passages": 14,
|
|
291
|
+
"pool_miss_cases": 1,
|
|
292
|
+
"pool_passages": 225,
|
|
293
|
+
"quality_eligible_cases": 14,
|
|
294
|
+
"unjudged_passages": 211
|
|
295
|
+
},
|
|
296
|
+
"metric_policy": {
|
|
297
|
+
"empty_positive_pool": "undefined",
|
|
298
|
+
"failed_eligible_response": "missing",
|
|
299
|
+
"gain": "2**native_grade-1",
|
|
300
|
+
"latency": "trusted_runner_process_wall_including_startup_and_model_loading",
|
|
301
|
+
"quality_population": "queries_with_positive_native_qrel_in_frozen_pool",
|
|
302
|
+
"relevance_threshold": 1,
|
|
303
|
+
"top_k": 5,
|
|
304
|
+
"unjudged_passage_effective_grade": 0,
|
|
305
|
+
"unjudged_passages_are_human_negatives": false
|
|
306
|
+
},
|
|
307
|
+
"provenance": {
|
|
308
|
+
"changes": [
|
|
309
|
+
"select previously opened 15 development queries",
|
|
310
|
+
"retain BM25 top15 order and scores",
|
|
311
|
+
"preserve absent native qrels as null with explicit effective-zero metric policy"
|
|
312
|
+
],
|
|
313
|
+
"distribution_rights_inferred": false,
|
|
314
|
+
"fresh_holdout": false,
|
|
315
|
+
"human_review_claimed": false,
|
|
316
|
+
"licenses": [
|
|
317
|
+
{
|
|
318
|
+
"asset": "claims_and_evidence",
|
|
319
|
+
"license": "CC-BY-4.0",
|
|
320
|
+
"source": "https://github.com/allenai/scifact/blob/master/LICENSE.md"
|
|
321
|
+
},
|
|
322
|
+
{
|
|
323
|
+
"asset": "abstracts",
|
|
324
|
+
"license": "ODC-BY-1.0",
|
|
325
|
+
"source": "https://github.com/allenai/scifact/blob/master/LICENSE.md"
|
|
326
|
+
},
|
|
327
|
+
{
|
|
328
|
+
"asset": "beir_repackaging_declared",
|
|
329
|
+
"license": "CC-BY-SA-4.0",
|
|
330
|
+
"source": "https://huggingface.co/datasets/BeIR/scifact"
|
|
331
|
+
}
|
|
332
|
+
],
|
|
333
|
+
"native_provenance_verified": true,
|
|
334
|
+
"native_qrel_source": "qrels/test.tsv",
|
|
335
|
+
"original_archive_sha256": "sha256:536e14446a0ba56ed1398ab1055f39fe852686ecad24a6306c80c490fa8e0165",
|
|
336
|
+
"original_archive_url": "https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/scifact.zip",
|
|
337
|
+
"previously_opened": true,
|
|
338
|
+
"review_receipts_verified": false,
|
|
339
|
+
"selection": "existing_scifact_development_15_bm25_top15",
|
|
340
|
+
"source_files": {
|
|
341
|
+
"corpus.jsonl": "sha256:dec31c8182f3d744c7d2c09423756fd1d17cbef75808db13ba01cc0aab4d1ac6",
|
|
342
|
+
"frozen-pools.json": "sha256:b3e6a31c62eccd90962bbce6372efe9bbf9fd397beedea7ea01d3a777e7d28f6",
|
|
343
|
+
"manifest.json": "sha256:739df50af296cd5760c5499e8711005f7bba35ad30e2aaa08d667fa19c45465d",
|
|
344
|
+
"qrels/test.tsv": "sha256:0864bb985e0ca2367ba217977e72004d549054b2b06666ed9d4825ac7c21284c",
|
|
345
|
+
"queries.jsonl": "sha256:8ff84a7c903f722981cd8d595c022660140c51867b27608a6d4910db86080313"
|
|
346
|
+
},
|
|
347
|
+
"use": "development_regression"
|
|
348
|
+
},
|
|
349
|
+
"registered_asset_locators": {
|
|
350
|
+
"beir.scifact.bm25-top15.dev15@1.0.0": "benchmark.json",
|
|
351
|
+
"beir.scifact.inputs.dev15@1.0.0": "inputs.json",
|
|
352
|
+
"beir.scifact.references.dev15@1.0.0": "references.json"
|
|
353
|
+
}
|
|
354
|
+
}
|