Initial experiment: soft optim vs normal

[jezgillen]

• Run the train loop
• Run inference to get e.g. 1000 inference samples (e.g. with proxy reward and prior reward)
• Get the proxy value cut-off using get_proxy_value_cutoff
• Run model twice: with soft-optim likelihood + tiny KL , and normal likelihood + tiny KL .
• Show the "breaks rules" stats for both runs