AF-HARM-01 / README.md
ClarusC64's picture
Upload 4 files
27ed5e6 verified
|
Raw History Blame Contribute Delete
3.86 kB
---
language:
- en
license: cc-by-4.0
pretty_name: "AIRFRAME — Decentralized Coherence & Relational Coupling Benchmark"
task_categories:
- question-answering
tags:
- airframe
- harmotence
- distributed-reasoning
- constraint-satisfaction
- relational-coupling
- synthetic-data
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train.csv
- split: test
path: data/test.csv
---
# AIRFRAME — Decentralized Coherence & Relational Coupling Benchmark
**Technical identifier:** AF-HARM-01
**Version:** 0.1 research preview
**400 challenges:** 320 train / 80 test, 80/20 per family.
AF-HARM-01 evaluates final-answer reasoning over local constraints in small distributed networks. It is part of the AIRFRAME / Harmotence research programme. These finite tasks do not establish a universal invariant or show that a model internally coordinates without a central controller.
## Challenge families
| Family | Required reasoning |
| --- | --- |
| `local_coherence` | Infer node phases from locally specified XOR relations and an anchor. |
| `coupling_adaptation` | Remove minimum-cost conflicting links to restore parity consistency, with deterministic tie-breaking. |
| `distributed_obstruction` | Select minimum buffer upgrades under local demand, capacity and transfer constraints. |
| `carrier_neutral_coordination` | Solve local phase constraints using arbitrarily relabeled carrier identifiers. |
The families use different constraint operations, not merely renamed shortest-path tasks. Local link information is supplied together in each prompt for reproducible offline scoring. This is a centralized *description* of a decentralized constraint problem, not a distributed execution protocol. No global clock is assumed.
## Data
Files: `data/train.csv`, `data/test.csv`. Columns: `id`, `family`, `challenge`, `expected_answer`, `answer_format`, `metadata`. Answers are JSON objects. Public test answers mean this is **not a blind benchmark**.
## Prediction format
Create `predictions.csv` with columns `id,prediction`. The `prediction` cell contains a JSON object with the exact keys specified in `answer_schema.json`. For example, for `local_coherence`, return `{"phases":{"N0":0,"N1":1}}` with all nodes from that challenge (the example is illustrative, not a gold answer). Use standard CSV quoting for embedded JSON.
The scorer compares exact JSON structure and values, ignores object-key order, preserves array order and distinguishes integers, floats and Booleans. A missing or malformed prediction is counted separately.
```bash
python scorer.py --predictions predictions.csv --gold data/test.csv
```
By default, output is `scores.json`, without gold answers or submitted predictions in the per-item details. `--include-answers` explicitly includes them. The scorer includes input and scorer SHA-256 hashes and per-family accuracy.
## Interpretation and limitations
- Tests finite local-constraint reasoning and exact output compliance, not internal model processes.
- No measured training improvement or carrier neutrality across model architectures is claimed.
- Node counts are small (5–8). Deterministic algorithms can solve the tasks; this is expected.
- The coupling-removal objective prioritizes total cost, then number of removals, then link-index order.
- The buffer-upgrade task allows each demand unit to be served locally or at an adjacent node; each directed neighbor transfer has capacity one.
- Public test answers can contaminate model comparisons. Do not use this split for private leaderboard claims.
## Attribution and source
Research programme: AIRFRAME / Harmotence. Dataset content under CC BY 4.0. The generator is private and is not part of the public release. `scorer.py` is provided for reproducible evaluation.