| { | |
| "model": "/content/rl_out/r1/hf", | |
| "max_new": 1024, | |
| "gsm8k_think": { | |
| "n": 500, | |
| "cap_rate": 5.8, | |
| "loop_rate": 5.2, | |
| "mean_len": 213.9, | |
| "secs": 77, | |
| "acc": 8.2 | |
| }, | |
| "math_think": { | |
| "n": 200, | |
| "cap_rate": 22.5, | |
| "loop_rate": 19.5, | |
| "mean_len": 387.7, | |
| "secs": 43, | |
| "acc": 5.0 | |
| }, | |
| "gsm8k_tool": { | |
| "n": 200, | |
| "cap_rate": 2.0, | |
| "loop_rate": 1.5, | |
| "mean_len": 87.5, | |
| "secs": 34, | |
| "acc": 8.0, | |
| "tool_use": 98.5, | |
| "tool_errors": 33 | |
| }, | |
| "gsm8k_plain": { | |
| "n": 200, | |
| "cap_rate": 1.0, | |
| "loop_rate": 1.0, | |
| "mean_len": 116.8, | |
| "secs": 31, | |
| "acc": 6.0 | |
| }, | |
| "chat": { | |
| "n": 200, | |
| "cap_rate": 10.0, | |
| "loop_rate": 8.5, | |
| "mean_len": 222.8, | |
| "secs": 47 | |
| } | |
| } |