d12 RL probe (15min, speed+convergence test)

RL probe run (speed + convergence data)

Source: SFT checkpoint (MLflow 2bffbfabb3024458ae2778a12f320163)

Goal: Run RL for ~15 min to measure:

• step time (how long per RL step on 1×H100)
• reward trend (is it converging?)
• extrapolate full epoch time and cost

Will be killed at 900s. This is intentional — just collecting data.

Status: funded_failed (timeout 900s)

Result: 0 training steps completed. Startup (tokenizer 60s + checkpoint 10s) + initial eval consumed entire budget. Cost: $3.34.

Fix: skip step-0 eval, longer max_time, or upload tokenizer to skip retrain.