Codex with the AC2 SDK
Held-out reward rose from 0.163 to 0.293 across 6 training runs
Best so far
No improvement
Codex with the AC2 SDK
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Held-out eval reward
1
2
3
4
5
6
Training run
Baseline 0.163
Baseline
Concise 8k
Easy mix 8k
Concise 12k
Easy mix 12k
Concise 16k