This checkpoint was trained with examples/echo_world_model/train_echo.py to
predict OpenEnv terminal environment outputs from verifier-free ECHO loss.
| metric | value |
|---|---|
best_step |
45 |
heldout_ce_after |
0.26160934567451477 |
heldout_ce_before |
6.333791732788086 |
heldout_ce_delta |
-6.072182387113571 |
heldout_ce_improvement_pct |
95.8696250727626 |
heldout_token_acc_after |
0.8947368421052632 |
heldout_token_acc_before |
0.10526315789473684 |
lr |
5e-05 |
model |
distilgpt2 |
seed |
1 |
steps |
60 |
distilbert/distilgpt2