val_bpb
1.1161
Architecture
Transformer
Optimizer
—
Artifact Size
~15.3 MB
Training Techniques
Quantization
GPTQ
bits: 6
scope: weights
Evaluation
sliding window eval
parameters: null
Test-Time Training
score-first TTT
parameters: null
Novel Contributions
- EGGROLL (Antithetic Ternary Bin Search)
- Post-GPTQ quantization refinement that optimizes INT6 bin assignments against BPB loss during evaluation budget
- Antithetic +1/-1 bin search on randomly selected quantized weight indices
- Adds missing eval_val_sliding_ttt call to the evaluation pipeline