Federal award outcome prediction tasks with recorded answers

Summary

Graded tasks built from public federal solicitations and the awards that followed them. Each task gives a model the solicitation, its amendments and the awarding office's history in that market up to a fixed cutoff, then asks for the award value band, the number of offers, the set-aside used or whether the incumbent kept the work. The answer is what the public record shows happened after the cutoff, and awardees appear as stable codes so models reason from history rather than recalled names. Built to order, with set size, agencies and difficulty mix agreed up front.

Collected to order. We collect it from the public record for your order. Scope, volume and timing are confirmed in writing before you sign.

Pricing on request

Request scope and pricing

At a glance

Planned scope. Volume, sources and timing are confirmed in writing before you sign.

Records: 48Kgraded tasks
Size: 1.2 GB
Tokens, estimated: 290M
Years: 2014 to 2026

Source: Public US federal solicitations, amendments and contract award records, paired by solicitation number and awarding office
Industry: Public sector
Data type: Evaluation and task packs
Region: United States
Languages: English
Formats: JSONL
Updated: 2026-10-14

Sources

System	Kind	Records
SAM.gov public contract opportunities	Public records	1.4M solicitations and amendments
Federal Procurement Data System public feeds	Public records	2.1M matched award actions

Good for

evaluating forecasting on events after a fixed information cutoff
measuring calibration on award value bands and offer counts
testing tool-using agents that research an office's buying history
tracking forecasting skill across quarterly refreshes with new awards

Fields

Field	Type	Description
task_id	string	Stable task id
task_type	enum	award_value_band, offer_count, set_aside_used or incumbent_retained
information_cutoff	date	Nothing dated after this appears in the context
context	object	Solicitation, amendments and the office's award history in that NAICS code
question	string	The question put to the model
answer_options	array<string>	Choices, for multiple-choice task types
answer	string	The recorded outcome
answer_source	string	The award action the answer comes from
grader	enum	exact_match, band_match or numeric_tolerance
difficulty	enum	easy, medium or hard, from baseline model accuracy

Example record

{
  "task_id": "T-52874-018342",
  "task_type": "award_value_band",
  "information_cutoff": "2025-04-30",
  "context": {
    "solicitation_id": "[SOLICITATION_211]",
    "title": "HVAC preventive maintenance and repair, [SITE_3] federal courthouse",
    "naics": "238220",
    "set_aside": "total_small_business",
    "office_history": "14 prior awards under this NAICS code at this office since FY2019, median 4 offers, incumbent [COMPANY_17]"
  },
  "question": "What will the base and all options value of the award be?",
  "answer_options": [
    "Under 250K USD",
    "250K to 1M USD",
    "1M to 5M USD",
    "Over 5M USD"
  ],
  "answer": "1M to 5M USD",
  "answer_source": "[ACTION_90417]",
  "grader": "exact_match",
  "difficulty": "medium"
}

One record. Placeholders such as [COMPANY_4] stay the same across the whole dataset.

Licenses available

Non-exclusive
Licensed to more than one buyer.

What ships with it

Datasheet: what it is, what it is good for, verified size, sources, years, format and limits.
Provenance file: the public sources, how they were collected and joined, and what was removed, mapped to the EU AI Act Article 53 training-data summary and California AB 2013.
Manifest with a hash for every file.
QC report, with what was removed or masked.
Your own copy, with unique canary records.

How it is prepared

Awardees, solicitation numbers and sites become stable codes, so models reason from history rather than recalled names. Personal contact details and the names of individuals in free text are removed before delivery.

How we protect data