---
pretty_name: REAM-288 activation-aware PLE n-gram selection
license: apache-2.0
task_categories:
  - text-generation
---

# REAM-288 activation-aware 50% PLE selection

This dataset contains the verified capacity-preserving row selection and remap
tables for compacting the PLE n-gram memory in
`sjakek/slimder-qwen38-ream288-depth32-agentic` to exactly 50% of its padded
physical capacity.

The selection combines a 250M-token multi-domain frequency census with a
3M-token proportional-stratified PLE activation-saliency pass. The saliency
sample contains 900k code, 750k agentic, 450k retrieval, 300k professional,
300k reasoning, 150k multilingual, and 150k structured tokens. Ranking keeps a
frequency-locked core, protects peak saliency rows, then fills the remaining
direct-row budget by activation-aware utility. Each dropped row maps to one of
65,536 fallback buckets per head.

## Main result

- Candidate/baseline direct-row overlap: 94.1634%
- Rows changed relative to frequency-only selection: 5.8366%
- PLE activation mass retained by frequency baseline: 94.0535%
- PLE activation mass retained by activation-aware candidate: 99.9942%
- Top 1% activation mass retained: 100%
- BF16 PLE storage: 102,400,491,520 to 51,200,231,680 bytes

The `activation-selection/` directory contains 16 direct-ID arrays and 16 full
remap arrays. `selection-validation.json` verifies hashes, uniqueness, ranges,
direct mappings, per-head capacities, and exact 50% padded physical capacity.
`ple-saliency/` contains the resumable raw sum/max/hit statistics used for the
ranking. `selection-comparison.json` contains per-head overlap, activation-mass,
and domain-frequency comparisons.

No model checkpoint, cache, credential, or unmodified source weight is included.