ctech-fle-gt / behavior_results.json
ctech123's picture
Duplicate from ctech123/Qwen3.8-Flash-Next-E384-FP4-V2
d07994b
Raw History Blame Contribute Delete
3.03 kB
{
"description": "Iteration-2 short-budget CVSS RL round; same 351 training prompts x 8 samples for both evaluations, temperature 1.0/top_p 0.95. iter2 file = it1 policy before the update; iter3 file = it2 policy after the update. Not held-out.",
"changes": {
"cvss_max_tokens": 2200,
"reward": "lenient score extraction + continuous absolute-error reward + exact bonus + cvss3 severity credit",
"training": "group by full (task,prompt), skip empty CVSS truncations from backpropagation",
"warm_start": "grpo/lora_it1/ckpt"
},
"it1_policy": {
"tasks": {
"cvss3": {
"samples": 880,
"mean_reward": 0.1678,
"avg_completion_tokens": 1998
},
"cvss4": {
"samples": 440,
"mean_reward": 0.0545,
"avg_completion_tokens": 2167
},
"emission": {
"samples": 360,
"mean_reward": 0.6089,
"avg_completion_tokens": 973
},
"format": {
"samples": 240,
"mean_reward": 1.0,
"avg_completion_tokens": 36
},
"guard": {
"samples": 128,
"mean_reward": 0.8047,
"avg_completion_tokens": 761
},
"math": {
"samples": 720,
"mean_reward": 0.7542,
"avg_completion_tokens": 333
},
"retire": {
"samples": 40,
"mean_reward": 0.66,
"avg_completion_tokens": 1040
}
},
"cvss_detail": {
"cvss3": {
"samples": 880,
"finished": 344,
"parsed": 279,
"exact": 16,
"within_0.3": 36,
"fake_zero": 31,
"exact_over_all": 0.0182,
"exact_over_finished": 0.0465,
"within_0.3_over_finished": 0.1047
},
"cvss4": {
"samples": 440,
"finished": 77,
"parsed": 39,
"exact": 2,
"within_0.3": 9,
"fake_zero": 0,
"exact_over_all": 0.0045,
"exact_over_finished": 0.026,
"within_0.3_over_finished": 0.1169
}
}
},
"it2_policy": {
"tasks": {
"cvss3": {
"samples": 880,
"mean_reward": 0.3527,
"avg_completion_tokens": 1783
},
"cvss4": {
"samples": 440,
"mean_reward": 0.2948,
"avg_completion_tokens": 1991
},
"emission": {
"samples": 360,
"mean_reward": 0.9169,
"avg_completion_tokens": 838
},
"format": {
"samples": 240,
"mean_reward": 1.0,
"avg_completion_tokens": 35
},
"guard": {
"samples": 128,
"mean_reward": 0.8125,
"avg_completion_tokens": 876
},
"math": {
"samples": 720,
"mean_reward": 0.8347,
"avg_completion_tokens": 348
},
"retire": {
"samples": 40,
"mean_reward": 0.64,
"avg_completion_tokens": 1019
}
},
"cvss_detail": {
"cvss3": {
"samples": 880,
"finished": 617,
"parsed": 553,
"exact": 28,
"within_0.3": 74,
"fake_zero": 26,
"exact_over_all": 0.0318,
"exact_over_finished": 0.0454,
"within_0.3_over_finished": 0.1199
},
"cvss4": {
"samples": 440,
"finished": 264,
"parsed": 201,
"exact": 11,
"within_0.3": 39,
"fake_zero": 1,
"exact_over_all": 0.025,
"exact_over_finished": 0.0417,
"within_0.3_over_finished": 0.1477
}
}
}
}