MichielBuisman/beap-data

BEAP Datasets — BRAK & Stroph-7

This repository contains the datasets for the Benchmark for Empirical Adaptability and Plasticity (BEAP). BEAP measures two distinct capabilities in language models:

• Adaptability (via the LES sub-benchmark): how quickly and efficiently a model can acquire new structured knowledge using LoRA fine-tuning across ranks 0–128.
• Plasticity (via the PHS sub-benchmark): how well a model retains prior learning while integrating new knowledge through full fine-tuning across six progressive difficulty buckets.

Both datasets exist in two evaluation modes: Generation (G), where the model must produce valid outputs, and Recognition (R), where it must identify the valid option in a multiple-choice setting.

---

Datasets

BRAK

BRAK is a synthetic programming language with a fixed grammar: Subject–Object–Verb sentence order, mandatory verb suffixes, and an asymmetric bracket system ()(, ][, }{) that encodes certainty levels through a rock-paper-scissors-style logic. Tasks are organised into six difficulty buckets:

Bucket	Focus
1	SOV statements, basic verb forms
2	Verb suffixes (ok, em, ith, vu, al)
3	Variable assignment (>>), conditional branching (iff)
4	Function definitions (func)
5	Pipelines (~~), anonymous functions (lam)
6	Tentative/imperative bracket semantics

Each bucket includes train, test, and cf_probe splits. The CF (catastrophic forgetting) probes are used in PHS to measure retention of earlier buckets after training on later ones. Recognition mode files follow the same naming convention with an _r_ infix (e.g. brak_r_b3_train.json).

File sizes per bucket:

Bucket	Train	Test	CF probe
1–2	500	100	50
3–4	750	150	50
5–6	1000	200	50

---

Stroph-7

Stroph-7 is a synthetic constrained poetry format. Each poem consists of three strophes of four lines each, subject to multiple simultaneous constraints: line word counts (7 / 5 / 7 / 5), an ordered digit sequence across the six long lines (1–6), per-strophe anchor words at line 4, strophe-initial alliteration, and the presence of at least one nature-lexicon word per strophe.

The dataset contains 1,000 training examples and 500 test examples in Generation mode. Recognition mode provides multiple-choice items where exactly one of four poems satisfies all constraints, with the three distractors each containing a different injected violation (word_count, digit_missing, digit_order, anchor_wrong, alliteration, or nature_absent).

---

Data Format

Generation mode items:

{
  "prompt": "Write a Stroph-7 poem.\n\nRules:\n...",
  "target": "<valid poem or BRAK program>",
  "params": { ... }
}

Recognition mode items:

{
  "prompt": "Below are four poems / programs ... Which is fully valid?",
  "correct_label": "B",
  "options": {
    "A": { "text": "...", "violation": "word_count" },
    "B": { "text": "...", "violation": null },
    ...
  }
}

---

Intended Use

These datasets are intended for use with the BEAP benchmark runner (beap.py). They are not natural-language corpora and are not suitable for pre-training or general language modelling. Their value is specifically in measuring structured rule acquisition and retention under fine-tuning.

---

Generation

All data is fully synthetic and deterministically reproducible. Fixed seeds are used throughout:

Split	Seed
Stroph-7 train	0
Stroph-7 test	10000
BRAK train	42
BRAK test	99000
BRAK CF probe	55000

To regenerate:

python beap.py --generate --data-dir ./beap_data

---

Citation

If you use these datasets, please cite:

@misc{beap2026,
  title  = {BEAP: Benchmark for Empirical Adaptability and Plasticity},
  author = {Buisman, Michiel},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MichielBuisman/beap-data}
}