PR #260
open[codex] Validate sliding-window post-quant evaluation on 1xH100 proxy
by Kevxn97View on GitHub
val_bpb
1.3276
Architecture
—
Optimizer
—
Artifact Size
—
Training Techniques
Evaluation
sliding window eval
parameters: {"stride":null,"context_length":null}
Other
other
Validate post-quantized evaluation on a 1xH100 proxy run and use it as a reproducible baseline for future search.
parameters: {"hardware":"1xH100","branch":"codex/sliding-window-eval-v1"}
Novel Contributions
- Adds sliding-window validation so scored tokens see near-max left context during evaluation
- Keeps the training path unchanged while modifying evaluation behavior
- Introduces configuration knobs for evaluation stride and batching
- Documents a local and RunPod workflow for reproducibility
- Records validated experiment details in the repo tracker
- Uses a real proxy run to validate final post-quantized performance