← Lab

Turn verifiable answers into rewards

CPU · Python 3.12.13 · 2.19 s · 2026-10-22 08:56:13 UTC+8

Saved run results

verifier_tests6
token_surrogate0.8500000238418579
sequence_surrogate0.75
Rewards:
 tensor([[0., 0., 1., 1.],
        [0., 0., 0., 0.]])
Advantages:
 tensor([[-0.9998, -0.9998,  0.9998,  0.9998],
        [ 0.0000,  0.0000,  0.0000,  0.0000]])

'Answer: 42' → 1.0
'Answer: 35' → 0.0
'Answer: 42 Answer: 35' → 0.0
'Answer: -3' → 1.0
'Answer: 42.5' → 0.0
'42' → 0.0

{'token_surrogate': 0.8500000238418579, 'sequence_surrogate': 0.75}

Mean selected-token log-probability change: [-0.13198477029800415, -0.08652293682098389, 0.11663806438446045, 0.07964777946472168]
Experiment output 
224
Optional exercise: fill answer and enable RUN_EXERCISE.
metrics.json
{
  "verifier_tests": 6,
  "token_surrogate": 0.8500000238418579,
  "sequence_surrogate": 0.75,
  "logprob_change": [
    -0.13198477029800415,
    -0.08652293682098389,
    0.11663806438446045,
    0.07964777946472168
  ]
}