nnai-bench / README.md
vynly's picture
Correct fresh-study link to the verified static Space host
1f59c69 verified
|
Raw History Blame Contribute Delete
7.44 kB
metadata
pretty_name: NNAI-Bench
language:
  - en
license: mit
task_categories:
  - question-answering
  - text-classification
tags:
  - non-neural
  - symbolic-ai
  - reasoning
  - synthetic
size_categories:
  - 1K<n<10K
configs:
  - config_name: scheduling_v02
    data_files:
      - split: train
        path: scheduling_v02/train.jsonl
      - split: validation
        path: scheduling_v02/validation.jsonl
      - split: test
        path: scheduling_v02/test.jsonl
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: validation
        path: data/validation.jsonl
      - split: test
        path: data/test.jsonl

NNAI-Bench v0.1

Separate neural comparison track, 2026-11-07: measured neural references use the same core questions with a separate leaderboard and schema-3 review route. Non-neural ranks and all frozen data remain unchanged. Protocol · measured evidence.

Measured additions, 2026-11-07: CAEL 0.3.9 (test-adapted) 71.00%, standalone Griot KenLM 29.78%, and the released count Bigram model 25.00% on all 900 items. The Bigram checkpoint tied all choices and always selected A. These are maintainer evaluations, not submissions or endorsements by third-party creators. Frozen evidence and reproduction details.

Model evaluations, 2026-10-19: measured 900-question CAEL and MICA Flame-W maintainer runs are published in model_runs/2026-10-19. MODEL_RUNS.md documents the frozen native interfaces and resources. They are not third-party community submissions or independently reproduced rows.

New, 2026-10-19: Scheduling v0.2 is a separate 300-case constraint-solving track (120 train / 60 validation / 120 test), with structured shift assignments, full-feasibility and coverage-adjusted partial scores, and per-constraint diagnostics. Select the scheduling_v02 config in the viewer, or download scheduling_v02/*.jsonl directly. The original core v0.1 data, manifests and scores remain unchanged. No blended v0.1/v0.2 score is computed.

Original, deterministic synthetic tasks for comparing non-neural AI systems. Leaderboard and submission form. Maintainer: vynly. Released 2026-10-18.

Dataset

Split Total Per task family
Train 600 100
Validation 300 50
Public test 900 150

Six families: arithmetic, deductive logic, ordering, state tracking, string operations, and table reasoning. Each has three complexity tiers (1–3). The test has 50 items per tier per family; train/validation have approximately equal tiers. All questions have four distinct options. Answers use A/B/C/D, in list order. Answer positions are balanced globally and shuffled independently of item IDs. Logic uses open-world semantics: missing evidence means unknown, not false. The option inconsistent is a distractor in v0.1, not a tested contradiction class.

Fields: id (string), split (string), category (string), difficulty (integer), prompt (string), choices (list of 4 strings), answer (A/B/C/D). Label-free inputs are in questions/. Complete generation programs and an independent answer checker are available in audit.jsonl and benchmark.py. The checker does not run submitted code. Source-specific SHA-256 hashes are pinned in manifest.json.

Run

git clone https://huggingface.co/datasets/vynly/nnai-bench
cd nnai-bench

Run your model offline against questions/test.jsonl, without reading answers or audit metadata. Write a JSON array with exactly one entry per test ID:

[{"id": "test-arithmetic-0000", "prediction": "A"}]

The real file must contain all 900 IDs. Predictions are A, B, C, D or ABSTAIN. example_predictions.json contains all IDs with ABSTAIN; it is a format template, not a baseline result. Python 3.10+; no pip dependencies are required to score:

python benchmark.py score predictions.json --questions data/test.jsonl --output report.json
python benchmark.py baselines
python benchmark.py generate

The generator reproduces the released question files using seed 20261018. The baselines command runs seeded random choice, training-label majority, and BM25 nearest-training-prompt retrieval. It creates predictions, reports and a local leaderboard.json. Runtime varies with hardware. Download and run code only after reviewing it.

Scoring and eligibility

Main score = unrounded arithmetic mean of per-family accuracies, displayed to two decimals. Abstentions count as incorrect; coverage and accuracy on answered items are also reported. All families have equal test sizes, so macro = micro here. Uniform random choice has expected accuracy 25%. No neural components are allowed at inference, including embeddings, rerankers, teacher APIs or remote helpers. Neural-derived training data is allowed only with a separate visible disclosure. Closed-book and offline-retrieval resource tracks are shown separately. Test-adapted entries must declare that status and can be filtered separately. See PROTOCOL.md for the full procedure.

Provenance and limitations

These examples are produced by original procedural generators, not scraped or copied from ARC, bAbI, HellaSwag, PIQA or another benchmark. The generator and documentation were created with coding-assistant help; no neural inference is needed to generate, score or run the three baseline controls. This provenance statement is about the benchmark assets, not a claim about submitted models.

This is a small, English-only, templated synthetic capability test. It is not a general-intelligence, language-fluency, factual-knowledge, perception, safety, or real-world performance benchmark. Shared templates make targeted hand-written solvers feasible. Exact prompt strings do not overlap between splits; this is not a guarantee of semantic independence or unseen-template generalization. All test labels, programs and seeds are public. The split is held out by convention only; there is no secret or contamination-proof test. Repeated test tuning must be declared. Future versions need new task families and independent community review. Benchmark bugs should be reported with an item ID.

No user data or third-party corpora are included. Original code and data in this benchmark are MIT-licensed. No CAEL/MICA model assets are relicensed or included.

Separate fresh-transfer study — 7 November 2026

A completed maintainer/AI-assisted 900-question native-interpretation study is published separately: results page and full evidence. The evaluated native candidate scored 777/900, with 123 abstentions counted wrong; new wording scored 327/450. Fixed neural single-letter readouts, known task families and a count-fallback activation difference limit comparison. See the configuration amendment. This is not a Qwen result or a replacement for the frozen core/scheduling scores. CAEL 0.3.9.1 retains the tested native task modules with declared release metadata/configuration differences.