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 |
40 |
heldout_ce_after |
0.27128568291664124 |
heldout_ce_before |
6.182467460632324 |
heldout_ce_delta |
-5.911181777715683 |
heldout_ce_improvement_pct |
95.61201600098846 |
heldout_token_acc_after |
0.8421052631578947 |
heldout_token_acc_before |
0.05263157894736842 |
lr |
5e-05 |
model |
distilgpt2 |
seed |
0 |
steps |
60 |
distilbert/distilgpt2