val_bpb
1.1247
Architecture
Transformer
Optimizer
—
Artifact Size
15,922,913 bytes
Training Techniques
Regularization
magnitude pruning
parameters: {"mode":"selective_prune_pre_quant","enabled":true}
Quantization
GPTQ
bits: 6
scope: all
Test-Time Training
TTT
parameters: {"enabled":false}
Evaluation
sliding window eval
parameters: {"exact":true}
Compression
lzma
level: null
Novel Contributions
- Publishes a non-record 16MB submission package for an existing clean prune-before-quant baseline
- Records-only packaging change with no new method or metric change
- Uses selective prune-before-quant boundary with no pre-quant TTT and no mixed export
- Provides validated packaged artifacts and exact terminal metrics for audit/reproduction