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 |
35 |
heldout_ce_after |
0.26466473937034607 |
heldout_ce_before |
6.246147632598877 |
heldout_ce_delta |
-5.981482893228531 |
heldout_ce_improvement_pct |
95.76275242056317 |
heldout_token_acc_after |
0.8421052631578947 |
heldout_token_acc_before |
0.10526315789473684 |
lr |
5e-05 |
model |
distilgpt2 |
seed |
2 |
steps |
60 |
distilbert/distilgpt2