CS 312 · Assignment 1
Experiment configurations and results.
Run ledger
Execution status
Model organism
The d8 path
One 1,024-token sequence flows through the same eight-block decoder used by every baseline run.
vocabulary 4,096hidden size 512RMSNorm → causal attention → + residualRMSNorm → SwiGLU MLP → + residual512 values / token4,096 logits / tokenmean cross-entropyCausal attention with rotary positions and query-key normalization.
Gate and up projections combine before the down projection.
1% warmup, linear decay, weight decay 0.1, gradient clip 1.0.
614.4M training tokens; batch size 64; held-out validation sequences.
Recorded measurements
Plot desk
0 runs in view
Terminal measurements
Hover or focus a point for its run record.Final validation loss by run order
Trajectories
Completed runs with recorded history.Training and validation loss
Problem views
Prescribed comparisons
Each panel reads directly from the same exported run records.
P1 · Hyperparameters
Terminal validation loss
P1 · Paired sweep
Cell labels are terminal validation losses.Learning rate × batch size
P2 · Scaling
Scaling families
P3 · Variation
One point per completed run, arranged by experiment group.Terminal-loss records
P4 · Amplification
Absolute paired loss gap
P6 · Internal statistics
Parameter, activation, and gradient RMS
Machine-readable evidence
Run records
| Run | Problem / group | State | Model | Configuration | Final train | Final validation | Runtime | Record |
|---|---|---|---|---|---|---|---|---|
| Loading run records… | ||||||||