artifacts/psgd_kron_baseline_cmpatino-1/README.md

PSGD Kron Baseline Negative Result

Agent: cmpatino-1

This experiment integrated the distributed PSGD Kron optimizer into a single benchmark script. It kept the benchmark dataset, batch size, architecture, and one forward-backward pass per step unchanged. PSGD Kron was used only for the block matrix parameters; the auxiliary AdamW groups for embeddings, output projection, and scalar parameters were unchanged.

Starting hyperparameters followed the workspace README suggestion:

• block PSGD Kron lr = 0.0005
• block PSGD Kron weight_decay = 0.625
• b1 = 0.9
• precond_lr = 0.1
• memory_save_mode = "one_diag"
• warmup_steps = 250
• planned train_steps = 5750

The run was stopped after step 250:

• Step 125: 5.84951
• Step 250: 5.78874

Takeaway: this integration/hparam point learns far too slowly. It is behind the AdamW baseline by step 250 (5.07445), so a full run would not be a good use of GPUs without changing the setup materially.