MarshallDoyle/nasa-ground-work-eval-v3

NASA Ground Work Eval: v3 development draft

This research bank asks whether models can produce correct scientific and engineering answers under a recorded evaluation protocol. Its intended use is to compare capability, cost, and specific failure modes in work relevant to NASA ground personnel. It is not an official NASA dataset, an operational mission model, or evidence of measured workforce benefit.

What is in this version?

There are 1,000 development instances across 20 selected science and engineering prerequisite topics:

Origin and intended collection	Instances	Labeled problem families
Attributed MMLU-Pro questions, challenge collection	600	600 source identities
New synthetic NASA-motivated work, natural-work collection	232	58
New synthetic NASA-motivated work, challenge collection	168	42
Total	1,000	700

Each new family has four related parameter cases. This is not 1,000 independently constructed questions. Source identities also do not guarantee that imported questions are scientifically independent. All difficulty labels are design intentions, not calibrated measurements. Many new tasks supply the governing equations explicitly; this draft does not establish extreme or frontier-level difficulty.

The 20 topics are astronomy, optics, classical mechanics, engineering thermodynamics, electromagnetics, electrical circuits, communications, fluid mechanics, heat transfer, transport phenomena, machine design, general chemistry, physical chemistry, biology, genetics, scientific mathematics, differential equations, statistics, computer science, and machine learning/security. Each has 30 imports and 20 synthetic instances. These are selected prerequisite strata, not a complete taxonomy of NASA's Earth science, planetary science, aeronautics, human research, or supporting occupations. Supporting-personnel workflows and executable repository repairs need further development; they are not established by a large question count.

The collection field keeps nasa_work and challenge separately scored within one dataset. The origin field separately identifies imports versus original synthetic cases. Report both axes; do not turn their mixture into one NASA-competence percentage.

The problem this benchmark is trying to solve

Generic examination benchmarks can test broad knowledge or reasoning without establishing whether a model produces a usable analysis, respects data provenance, handles uncertainty, or reduces the effort required to check its work. Our earlier v1 pilot illustrated this gap: 90 of 108 questions passed for all 11 tested models. High scores on those development questions were weak evidence for model selection, particularly when the remaining failures included JSON formatting and API problems.

The intended distinction here is to request inspectable work products: corrected measurements, uncertainty calculations, resource estimates, bounded engineering decisions, dataset-quality operations, and reproducible intermediate quantities. The new tasks include correlated uncertainty, calibration publication dates, missing-value/quality-bit semantics, numerical stability, conservation, false-positive base rates, and competing constraints. Values and experimental situations are synthetic.

NASA's Evaluation Policy, NPD1210.7 describes rigor, relevance, utility, independence and transparency. NASA's 2024 advisory-council AI briefing provides a historical management perspective on useful agency applications. The NASA technical memorandum on LLM assurance arguments is research, not agency leadership policy. Our task design is an interpretation of these sources, not NASA's approval or a claim that these calculations operationalize every agency requirement.

Independent expert review, realistic files and workflows, a genuinely protected holdout, and human-only versus AI-assisted studies remain necessary to establish scientific validity and workplace value.

How this relates to existing benchmarks

Benchmark	What we use or learn	What we do not claim
MMLU-Pro	600 pinned, attributed source questions; multiple-choice foundation coverage	A paraphrase becomes a new independent question, or a familiar source is uncontaminated
GPQA	Expert science reasoning and independent review are useful design principles	We downloaded, rewrote, or redistributed GPQA examples; its official gate asks users not to reveal examples online
SciCode	Eligible source bank for a future executable scientific-coding collection	Downloaded question/test files alone establish a verified, runnable grader; no SciCode tasks are scored in v3 yet
AA-LCR	Long-document evidence handling and versioned reference corrections	Its question license automatically licenses all underlying source documents
AstaBench	Real scientific work, explicit agent protocols, cost/performance comparisons	Our short numerical cases are equivalent to end-to-end research projects

MMLU-Pro's dataset is declared MIT by its card and paper; its code repository has a separate Apache-2.0 license. Original source labels, revision, row identifiers and notices are retained. MMLU-Pro's authors describe the underlying STEM website material as open-licensed, but this project could not independently retrieve that website's original license. This limitation is recorded in the source register and attribution notice. Source answers have not undergone independent domain-expert adjudication here. Rewording would not resolve either uncertain answer quality or training contamination. This release preserves source wording so that changes and origins remain auditable.

No license has been selected for the user's original research. This repository begins as a private development draft. Source material keeps its respective applicable terms. The source license must not be treated as a blanket license for the entire collection.

Examples and answer structure

Imported MMLU-Pro astronomy item 9364 (nasa-v3-mmlupro-9364) asks for the best approximation to the Sun's surface temperature. Its 10 supplied options include 6000 K; the inherited correct label is C in the stored option order. This is a foundation check, not evidence of frontier-level difficulty. The original dataset, item ID, revision and option order accompany the stored row.

The new data-processing task nasa_v3_computer_science_quality_mask_00 supplies values [10, -9999, 20, 30, 40], flags [0, 0, 1, 2, 0], and weights [1, 2, 3, 4, 2]. It defines -9999 as missing, bit 0 as invalid, and bit 1 as advisory. Correct processing retains zero-based indices 0, 3, and 4, with weight sum 7 and weighted mean (10×1 + 30×4 + 40×2)/7 = 30. The requested work product is:

{
  "retained_count": 3,
  "retained_weight_sum": 7,
  "weighted_mean": 30.0,
  "retained_index_sum": 7
}

This tests a common failure: deleting advisory-only data or allowing fill values into a weighted mean. The expected values follow the supplied synthetic semantics; they do not assert a particular NASA product's quality conventions.

The uncertainty task nasa_v3_statistics_correlated_replicates_00 specifies 20 measurements with individual standard deviation 2 and pairwise correlation 0.1. It asks for the mean's variance, standard error, effective sample size, and the naive independent-measurement standard error. The correct values are approximately 0.58, 0.7615773, 6.8965517, and 0.4472136, respectively. The variance is (2²/20)×[1 + (20−1)×0.1] = 0.58; the other quantities follow from its square root, 20/[1 + 19×0.1], and 2/√20. This tests whether repeated correlated observations are incorrectly treated as independent evidence.

Scoring

The imported items have between 3 and 10 choices (548 of 600 have 10). Uniform random guessing has an expected accuracy of approximately 10.76% on this selected import bank. This baseline does not apply to multi-field numerical tasks. See answer-structure audit.

• Multiple choice: compare the selected label to the inherited key. A bare correct label is recorded as correct content but invalid JSON format.
• Numerical work: check every requested quantity against the stated reference calculation and task-specific tolerance. New floating-point values generally allow relative error 0.1%; small absolute allowances are capped to avoid accepting zero for tiny nonzero quantities. Integer counts and Boolean decisions are exact. Inspect each row for its actual tolerance.
• Full task success: all critical answer checks plus the requested response format must pass. Partial credit reports the fraction of weighted checks passed.
• Keep scientific-content errors, format errors, incomplete responses, transport errors and missing responses separate.
• These checks do not establish the truth of unrequested prose, citation entailment, or an expert's acceptance of a complete deliverable. human_review_required remains true and fully_accepted remains unknown.

The reference-output acceptance and mutation tests validate evaluator behavior. They do not independently prove that every reference answer is scientifically correct. See the mechanical validation and scientific review records.

The content parser can also recover one unambiguous JSON object embedded in prose or a Markdown fence. Such an answer can receive content credit while still failing the requested format. Multiple candidate objects and duplicate keys are rejected. This parser change is recorded in the pilot's rescoring provenance; original responses and original run scores are retained.

Storage and access

The dataset is physically stored as UTF-8 JSON Lines on the Hugging Face Hub. Each line is one record.

• hub/dev.jsonl: the standard Hub view; variable nested candidate, scoring and provenance objects are JSON strings for reliable Arrow loading.
• data/dev.jsonl: canonical nested records for the local evaluation harness.
• candidate_packets/dev.jsonl: candidate-only projection plus task identifiers for transport. The harness itself sends only the approved prompt/context/choices/schema.
• evaluation/evaluation.py: standalone deterministic scoring code.
• authoring/: construction-source snapshots and provenance, not an independent human review.
• source_notices/: attribution and applicable original source notices.
• dataset_manifest.json: counts, source status, hash, and limitations.

After signing in to an account with access, load the Hub view with:

from datasets import load_dataset
import json

ds = load_dataset("MarshallDoyle/nasa-ground-work-eval-v3", split="dev")
candidate = json.loads(ds[0]["candidate_json"])
scoring = json.loads(ds[0]["scoring_json"])

Never mount this full answer-containing development repository into an evaluated agent's filesystem or retrieval tools. Private Hub visibility does not make it a holdout.

Model comparison protocol

The first v3 pilot uses a fixed, seeded, family-distinct subset with matched questions for every model, no retries and an explicit dollar ceiling. Record exact model/provider, reasoning setting, output cap, protocol, dataset hash, raw response, usage, cost and status. The planned low-cost panel includes OpenAI, Anthropic, Google, DeepSeek and Qwen. A budget-conscious pilot measures those settings; it is not a price-agnostic maximum-capability leaderboard.

No-tool requests supply no tools, specify tool_choice: none, and disable web and response-healing plugins. This is request-level control. Provider network isolation and locked account plugin overrides are not independently verified. Search-enabled comparisons need a separately fixed protocol and full tool-cost accounting; they have not been run in this release.

Answer-containing public benchmark sources may be searchable on the web. A future web-enabled score on those items would mix answer retrieval with reasoning; it must not be presented as a pure reasoning comparison. A realistic scientific-work condition should also distinguish permission to use a calculator or code from permission to search the web.

Reports should show matched-case score versus observed cost, per-topic pass/count cells, and failure types. Group related variants together when sampling or splitting. Do not use row-level binomial confidence intervals as if four variants were independent expert-authored problems. A 40-task pilot provides at most two observations per topic, which is insufficient for robust topic rankings.

Results: initial development pilot

The October 8, 2026 baseline made 195 attempted calls across five models. Provider interruptions left 38 questions attempted by every model: 19 imported multiple-choice questions and 19 original synthetic tasks. Five attempts from two partly completed question blocks are preserved but excluded from the matched comparison. No failed or refused baseline answer was retried or silently replaced. All complete responses were rescored consistently under the documented content extractor.

For the same 19 original synthetic tasks, with no external tools and an output cap of 1,536 tokens:

Model	Delivered correct answers / attempted	Correct / evaluable answers	Reported API cost for these 19 attempts
Claude Sonnet 5.5	19/19	19/19	$0.098810
GPT-6 Luna	18/19	18/19	$0.003519
Gemini 3.8 Flash	15/19	15/19	$0.041061
DeepSeek V4.1 Flash	10/19	10/10	$0.020992
Qwen3.8 Flash	8/19	8/9	$0.006080, plus at most $0.021375 in unresolved charges

DeepSeek's nine missing answers in this subset hit the output limit. Qwen had eight rate-limit errors and two returned fragments with no evaluable answer. Conditional accuracy therefore does not measure dependable answer delivery. Extra prose or Markdown around otherwise correct JSON is a format failure, not automatically a scientific error. These are single attempts on a small development subset, with low reasoning settings (Qwen used its supported explicit reasoning-token setting); they do not establish the best model at maximum capability or a robust domain ranking.

A separate output-budget condition evaluated all 19 of these synthetic tasks again on DeepSeek with a 6,144-token cap, holding the other configured inference settings fixed. All 19 paired requests used the same provider (Relace), and their payloads differed only in the output cap. It delivered 16/19 correct answers, with three length-limited responses, for $0.047542. Every completed answer passed the declared content checks. This is a separately reported experimental condition; it does not replace the baseline, and one generation per condition leaves stochastic variability uncontrolled.

Across the baseline and this additional condition, reported new charges were $0.3282547365, with $0.08540294 of conservatively retained unresolved charges. The combined upper accounting bound is $0.4136576765, within the previously authorized campaign budget. Exact receipts, settings, source hashes, uncertainty about charges and comparison data accompany the result package.

Answer-quality audit

The baseline contained eight evaluable disagreements with reference keys. Six were original numerical answers; separate high-precision recalculation confirmed the references under the supplied assumptions. Two involved imported multiple-choice questions: diode item12125 omits the intended device-model assumptions; thermal-panel item11407 omits conventions required to reproduce its exact numerical key. Both need expert adjudication. This does not establish that the models' alternatives are correct. The original inherited-key scores are retained, and a separately labeled, post-hoc sensitivity analysis excludes both flagged items for every model.

The pilot samples only one of the newly authored challenge families. It therefore does not calibrate that collection's intended difficulty. Sonnet also passed all 19 synthetic tasks in the tested subset: we have not demonstrated the extremely difficult frontier benchmark envisioned for a later release. More source imports or parameter variants alone will not close that gap.

See portable baseline responses, matched comparison summary, and failed-answer review. Plotting scripts, editable data and rendered figures are included in the plotting package.

Download the editable Matplotlib/Spyder package. Open plot_v3_results.py for the model comparison or plot_v3_cap_study.py for the output-budget experiment. The package includes local data, PNG/SVG exports, CSV tables and instructions. The scripts were executed and the rendered figures inspected; Spyder's graphical application itself was not tested.

Development status

The row-level evaluation_status records the frozen pre-pilot authoring snapshot. Subsequent attempts and results belong in separate versioned result files; the evaluated questions and answers are not silently rewritten after seeing model behavior. The NASA application map records our proposed relevance and the remaining scope gaps.

All 1,000 reference responses pass the declared evaluator. A mechanical audit rejects single-field wrong-answer mutations, confirms imported answer mapping and verifies candidate projection. Assistant cross-review checks formulas and invariants, while independent human expert review remains pending. The original 400 instances were authored in this assistant session; no external question-authoring API was used and the session's exact model identity is not independently verified.

The next validity step is expert adjudication and pilot-error review, followed by genuinely new task families involving real public scientific data and executable work. Do not silently remove questions simply because a favored model fails them. Record reference corrections and score all affected models again on a new version.