val_bpb
1.0607
Architecture
Transformer
Optimizer
—
Artifact Size
Under 16MB
Training Techniques
Test-Time Training
full TTT
parameters: {"phases":4,"docs_per_phase":2500,"total_doc_evals":10000}
Novel Contributions
- 4-phase test-time training adaptation pass
- 2500 documents per phase
- 10,000 total TTT document evaluations
- Local-to-official evaluator BPB offset correction