AF-RECOND-01 / README.md
ClarusC64's picture
Upload 4 files
a49f37d verified
|
Raw History Blame Contribute Delete
3.58 kB
---
language:
- en
license: cc-by-4.0
task_categories:
- question-answering
pretty_name: "AIRFRAME — Obstruction & Constraint Revision Benchmark"
tags:
- airframe
- apertience
- relational-reasoning
- constraint-revision
- synthetic-data
- reasoning-evaluation
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train.csv
- split: test
path: data/test.csv
---
# AIRFRAME — Obstruction & Constraint Revision Benchmark
**AF-RECOND-01 v0.1** is a synthetic reasoning dataset with 400 generated challenges (320 train / 80 test). It asks systems to respond to changed constraints rather than repeat an earlier solution.
## Four challenge families
| Family | Records | Task |
|---|---:|---|
| Constraint invalidation | 100 | Recompute the least-cost directed route after an edge disappears |
| Minimal reconditioning | 100 | Choose an option and the least costly necessary requirement relaxations |
| Compound obstruction | 100 | Choose a feasible pair under joint resource, duration, and exclusion constraints |
| Nonclosure recovery | 100 | Count and identify alternative simple paths after node exclusion |
Each family has 80 training and 20 test instances. The data uses bounded finite problems with exact answers. It does **not** establish internal frame release, general agentic recovery, or novel intelligence mechanisms.
## Files
- `data/train.csv` — public training examples with gold answers
- `data/test.csv` — public test examples with gold answers, **not a blind benchmark**
- `scorer.py` — deterministic final-answer evaluator
- `generate.py` — deterministic generator, seed 20261115
## Columns
`id`, `family`, `challenge`, `expected_answer`, `initial_state`, `obstruction`, `instance`, `scoring_rule`.
The `challenge` contains the task and input instance. `expected_answer` is a JSON object serialized as a CSV string. `instance` contains machine-readable source parameters. `initial_state` may be `null` for families without a single prior solution. `obstruction` identifies the constraint being revised or imposed.
## Run the scorer
Create `predictions.csv` with columns `id,prediction`. The prediction cell must contain a valid JSON object matching the requested output shape. Use CSV quoting around JSON values.
```bash
python scorer.py --predictions predictions.csv --gold data/test.csv
```
The scorer writes `scores.json` with total and per-family accuracy, missing and invalid-format counts, and SHA-256 hashes. It does not disclose gold answers in its output unless `--include-answers` is passed. Note that the public CSV already contains answers.
## Interpretation and limits
A correct response demonstrates final-answer accuracy on the specific finite problem. It does not prove the model detected an obstruction internally or performed the intended reasoning. A programmatic solver can solve these tasks. The train/test split is stratified, but examples share generator families and may admit heuristics. Do not use this release to claim superiority over existing reasoning benchmarks or to infer internal representations.
## Reproducibility
Run `python generate.py` to recreate the two CSV splits. The generator uses only Python's standard library. The dataset is offered under CC BY 4.0; source-code files can be separately licensed under MIT if their copyright holder chooses.
## Citation
AIRFRAME, *AF-RECOND-01: Obstruction & Constraint Revision Benchmark*, version 0.1 (2026). Cite the dataset repository URL and version used.