PR #475
open[Non-record] ABRAM_CHIP v2 — HECR int16 ultra compact — 34 KB — 0.50 bpb
by abrahaw123-cellView on GitHub
val_bpb
0.5000
Architecture
—
Optimizer
—
Artifact Size
34 KB
Training Techniques
Quantization
int16
bits: 16
scope: all
Architecture
HECR quantum state vectors
Novel architecture using HECR quantum state vectors.
parameters: null
Other
other
No transformers, no gradients, and no floats are used.
parameters: null
Novel Contributions
- Novel architecture using HECR quantum state vectors
- int16-only model with no transformers, no gradients, and no floats
- Ultra-compact 34 KB model size
- Reported val_bpb of approximately 0.50 on a simulation dataset