Dataset Viewer
Auto-converted to Parquet Duplicate
performer_id
stringlengths
14
16
training_type
stringclasses
4 values
completion_date
int64
15
1.17k
⌀
perfor_0000000
one-off
589
perfor_0000001
standard
641
perfor_0000002
expedited
690
perfor_0000003
expedited
81
perfor_0000004
standard
606
perfor_0000005
expedited
796
perfor_0000006
one-off
705
perfor_0000007
recurring
233
perfor_0000008
one-off
97
perfor_0000009
standard
367
perfor_0000010
standard
602
perfor_0000011
recurring
948
perfor_0000012
standard
479
perfor_0000013
expedited
1,025
perfor_0000014
standard
622
perfor_0000015
one-off
758
perfor_0000016
standard
573
perfor_0000017
expedited
483
perfor_0000018
standard
268
perfor_0000019
recurring
380
perfor_0000020
standard
1,029
perfor_0000021
one-off
931
perfor_0000022
standard
1,139
perfor_0000023
recurring
508
perfor_0000024
standard
268
perfor_0000025
recurring
208
performer-id-231
one-off
26
perfor_0000027
expedited
161
perfor_0000028
standard
113
perfor_0000029
one-off
251
perfor_0000030
standard
357
perfor_0000031
standard
1,154
perfor_0000032
expedited
587
perfor_0000033
standard
366
perfor_0000034
standard
894
performer-id-035
standard
267
perfor_0000036
expedited
55
perfor_0000037
one-off
497
perfor_0000038
expedited
738
perfor_0000039
expedited
365
perfor_0000040
expedited
173
perfor_0000041
expedited
325
perfor_0000042
one-off
368
perfor_0000043
one-off
59
perfor_0000044
one-off
655
perfor_0000045
recurring
819
perfor_0000046
expedited
131
performer-id-584
one-off
718
perfor_0000048
standard
57
perfor_0000049
one-off
283
perfor_0000050
one-off
412
perfor_0000051
standard
780
perfor_0000052
recurring
188
perfor_0000053
recurring
913
perfor_0000054
expedited
90
perfor_0000055
standard
527
perfor_0000056
standard
1,171
perfor_0000057
recurring
461
perfor_0000058
one-off
212
perfor_0000059
standard
741
perfor_0000060
expedited
668
perfor_0000061
expedited
173
perfor_0000062
one-off
772
perfor_0000063
one-off
823
perfor_0000064
recurring
186
perfor_0000065
recurring
607
perfor_0000066
expedited
515
perfor_0000067
standard
1,028
perfor_0000068
recurring
513
perfor_0000069
standard
1,122
perfor_0000070
expedited
784
perfor_0000071
one-off
868
perfor_0000072
expedited
1,108
perfor_0000073
expedited
1,127
performer-id-828
standard
308
perfor_0000075
standard
1,152
performer-id-717
expedited
611
performer-id-076
one-off
270
perfor_0000078
one-off
793
perfor_0000079
one-off
894
perfor_0000080
standard
922
perfor_0000081
expedited
854
perfor_0000082
standard
292
perfor_0000083
standard
915
performer-id-058
standard
605
perfor_0000085
recurring
761
perfor_0000086
expedited
356
perfor_0000087
recurring
1,096
perfor_0000088
recurring
110
perfor_0000089
expedited
559
perfor_0000090
standard
521
perfor_0000091
one-off
550
perfor_0000092
expedited
1,137
perfor_0000093
standard
758
perfor_0000094
standard
1,001
perfor_0000095
recurring
726
perfor_0000096
standard
296
perfor_0000097
expedited
1,036
perfor_0000098
recurring
979
perfor_0000099
expedited
587
End of preview. Expand in Data Studio

agent-eval-profiles

Evaluation profiles for Orama Studio's agents. Private. Internal-confidential.

A profile is a situation plus its data: who someone is, what they are responsible for, what data they can reach, and the actual schemas and records. It is a tool for evaluating models, not a fixture for an engine — everything here is scoreable without a running Zetta node.

What it evaluates

profile  ──►  Auto-Suggest model  ──►  a set of candidate TASKS
                                            │
profile + chosen intent  ──►  Project-creation agent  ──►  a Pipeline / Project
  • Auto-Suggest is scored on whether the tasks it proposes are the right set for that situation. This is why the role and its constraints are part of the profile: the same data has different correct answers for a fleet director who can dispatch 40 jobs a month and one who can dispatch 400.
  • Project creation takes the profile plus a chosen intent and is scored on whether the Pipeline it builds is correct and production-grade for that data.

Layout

<profile>/situation.md      the person, their responsibilities, their connections
<profile>/schema.json       tables, row counts, columns
<profile>/records/*.jsonl   the actual records

The rule that makes these worth running

Awkwardness is generated, never annotated. If subscriptions are paused rather than cancelled, the records really are like that — the profile does not say so. Whether a model notices is exactly what is being measured.

The quality bar: perturb one fact and the correct answer must change. If it does not, the profile tests recall of the ask rather than reading of the data.

Profiles

17 profiles · 1.69M records. Four catalogue solutions × four personas, plus the original hand-built one.

Variation comes from situations over real schemas, not from inventing tables. Random noise shaped like a schema teaches nothing, so each solution contributes one schema family with genuine structure, and the personas vary what actually changes the correct answer: authority, capacity, obligation, cadence. The same tables under a different situation are a different eval — that is the point rather than a shortcut.

Solution Slug The structure a model must read
S-1 churn churn-* The label lives in accounts.churn_date, not the events table — the join is the job. comm_events.sfdc_account_id is null on 11.4% of rows, and event_updated_datetime_pt is the pipeline load date rather than the source change date.
S-8 predictive maintenance predictive-maintenance-* work_orders.closed_at is null on 31% of rows. Using it as a label window silently drops every currently-failing asset.
S-12 claims triage claims-triage-* total_paid is only populated after closed_at. Training on it leaks the outcome into a decision made at FNOL. The honest target is siu_referred, 2.3% positive.
S-5 demand forecasting demand-forecasting-* Subscriptions largely determine daily orders. A model that reads the subscription table predicts most of orders_daily from cadence alone; one that forecasts the aggregate produces a backtest a spreadsheet beats.
Persona The constraint that changes the answer
Director of Operations Can action ~40 items a month and no more.
Senior Analyst Anything shipped must be reproducible by a colleague from the query alone.
VP, regulated line Must explain any individual decision; no decision without a named human approving that instance.
Head of Function Orders placed weekly on a fixed day, unchangeable once placed.

None of this is stated in the data. Whether a model notices is what is being measured.

Limitations — read before using

  • Four schema families, not twelve. The catalogue has twelve Band-A modules; eight have no profile yet. Generating shallow ones to reach a number would violate the rule that makes these worth running.
  • The personas are archetypes, not researched roles. They vary the right axis (what changes the answer) but were not drawn from real job descriptions.
  • Nothing has been scored against these yet. No model has been evaluated, so there is no baseline and no evidence the profiles discriminate. Until then this is a dataset, not a benchmark.
  • Synthetic. The structure is deliberate and the schemas are plausible, but no profile is drawn from a real customer's data.
Downloads last month
8