8Planetterraforming/parameter_golf_v14_quecto_recovery_tower_english

Parameter-Golf-V14-Quecto-Recovery-Tower-English

English-only auxiliary dataset for compact-model and BPB-oriented experiments.

V14 centers on the compression tower idea: store reusable structure rather than every byte flatly. It also adds the write-delete-recover distinction: training can distill structure into weights, codebooks, caches, residuals, or generators, but exact recovery is possible only when the required information is stored or generated by a reversible rule.

The dataset is synthetic and sanitized. It does not contain real personal data, real incident evidence, private files, or validation text.

Task families

• —compressiontowerv14: honest 1 TB / 1 GB / 100 MB to 16 MB reasoning.
• —recoverablestatev14: write-delete-recover, exact vs predictive recovery, reversible state, residuals, checksums.
• —privacywebfiltering_v14: fast signal extraction, data minimization, web-noise skipping.
• —alphabetmappingv14: strict 1-26 to A-Z mapping, no duplicates, no gaps, no out-of-range values.
• —siprefixv14: pico, atto, ronto, quecto, and naming discipline.
• —parametergolfbpb_v14: legal ablation planning under artifact and evaluation constraints.
• —promptmemorybudget_v14: compact execution order for complex prompts.

Suggested use

Use only as a small auxiliary dataset for experiments. Start with a 0.1% to 0.5% token mix against the main corpus. Test 1.0% only if the smaller mix is stable. Do not replace the main corpus.

Important limit

This dataset does not prove that arbitrary 1 TB can be losslessly compressed into 16 MB. It teaches the model the correct distinction: arbitrary lossless compression is impossible at that ratio, but structured sources can be represented by compact rules, generators, codebooks, residuals, and checksums.