HDC-AM · Master Knowledge Folding

Hyperdimensional Computing — Address Mapping. Stream HF master corpora → fold bipolar {-1,+1}4096 fast-weights → 444.4 MiB overlay.
D = 4096 bipolar {-1,+1}^D 113,766 slots × int8 = 444.4 MiB mmap zero-copy < 10 ms TR-1 strict ±1

Live demo — Hamming overlap & noise margin (browser-side, D=256)

query:
bit-flip noise 0%
slot rows: D=256 int8 bipolar in wasm/JS.
…

Budget math

itemvalue
dimension D4096
alphabet{-1, +1}D (int8)
slots (rows)113,766
bytes / slot4096 B
total465,987,174 B = 444.4 MiB
streaming loadersFineWeb-Edu · English Wikipedia · ArXiv (RAM clamp, no raw text on disk)

Validation gates before release

checkthresholdstatus (master run)
mmap zero-copy open< 10 msPASS
TR-1 strict bipolarityevery slot row ∈ {-1,+1}^DPASS
associative recalltop-1 query marginPASS

Artifact & engine

fast_weights_master.safetensors
444.4 MiB overlay — generated by a PRO-sandbox run of this pipeline (HF Spaces now require PRO for Gradio runtimes). Local recipe:

commandnotes
python hdc_am_pipeline.py ingest --d 4096 --slots 113766 --chunk-docs 512 --filename fast_weights_master.safetensorsfile corpus
python hdc_am_pipeline.py stats --overlay fast_weights_master.safetensorsmmap + TR-1
python hdc_am_pipeline.py recall --overlay fast_weights_master.safetensors --queries "Taj Mahal" --topk 8associative recall

Engine source (hosted here): hdc_am_pipeline.py · ingest_streamer.py · requirements.txt · run_master.sh · README.md