The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
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.
- Downloads last month
- 983