gemma-4-31B-it-abliterated-v3 / eval /optimization_summary_trial40.json
simonko912's picture
Duplicate from wangzhang/gemma-4-31B-it-abliterated
da378b3
{
"generated_at_utc": "2026-04-29T05:00:40.396149+00:00",
"base_model": "google/gemma-4-31B-it",
"repo_id": "wangzhang/gemma-4-31B-it-abliterated",
"selected_trial": {
"rank": 1,
"trial_index": 40,
"trial_number": 39,
"refusals": 7,
"kl_divergence": 7.324366038119479e-07,
"length_deviation": 0.0008759591017432696,
"vector_index": null,
"parameters": {
"attn.k_proj": {
"max_weight": 5.733064678080953,
"max_weight_position": 36.22021408860319,
"min_weight": 0.30663881606010823,
"min_weight_distance": 6.349510997322486
},
"attn.o_proj": {
"max_weight": 3.7950120964128624,
"max_weight_position": 37.753897993063156,
"min_weight": 2.596066328790265,
"min_weight_distance": 19.778650038422843
},
"attn.q_proj": {
"max_weight": 5.671821506957715,
"max_weight_position": 41.381947330490405,
"min_weight": 0.019698308484936443,
"min_weight_distance": 12.481719729578046
},
"attn.v_proj": {
"max_weight": 1.4212398130247585,
"max_weight_position": 48.19212212356811,
"min_weight": 0.7177214439809124,
"min_weight_distance": 9.55056279098482
}
}
},
"top3_trials": [
{
"rank": 1,
"trial_index": 40,
"trial_number": 39,
"refusals": 7,
"kl_divergence": 7.324366038119479e-07,
"length_deviation": 0.0008759591017432696,
"vector_index": null,
"parameters": {
"attn.k_proj": {
"max_weight": 5.733064678080953,
"max_weight_position": 36.22021408860319,
"min_weight": 0.30663881606010823,
"min_weight_distance": 6.349510997322486
},
"attn.o_proj": {
"max_weight": 3.7950120964128624,
"max_weight_position": 37.753897993063156,
"min_weight": 2.596066328790265,
"min_weight_distance": 19.778650038422843
},
"attn.q_proj": {
"max_weight": 5.671821506957715,
"max_weight_position": 41.381947330490405,
"min_weight": 0.019698308484936443,
"min_weight_distance": 12.481719729578046
},
"attn.v_proj": {
"max_weight": 1.4212398130247585,
"max_weight_position": 48.19212212356811,
"min_weight": 0.7177214439809124,
"min_weight_distance": 9.55056279098482
}
}
},
{
"rank": 2,
"trial_index": 46,
"trial_number": 45,
"refusals": 9,
"kl_divergence": 6.601356972169015e-07,
"length_deviation": 0.00262787730523012,
"vector_index": null,
"parameters": {
"attn.k_proj": {
"max_weight": 5.851557283702162,
"max_weight_position": 43.159126005754494,
"min_weight": 1.0554409058407976,
"min_weight_distance": 1.5398354922133537
},
"attn.o_proj": {
"max_weight": 4.575081848270157,
"max_weight_position": 41.9673193895106,
"min_weight": 1.8069519625068189,
"min_weight_distance": 22.008988042159274
},
"attn.q_proj": {
"max_weight": 5.1037476328037865,
"max_weight_position": 39.338308826072065,
"min_weight": 0.014776692029917217,
"min_weight_distance": 13.349683086944012
},
"attn.v_proj": {
"max_weight": 1.8622849844581164,
"max_weight_position": 42.95698075509393,
"min_weight": 0.7390747250972998,
"min_weight_distance": 11.011794606001594
}
}
},
{
"rank": 3,
"trial_index": 53,
"trial_number": 52,
"refusals": 12,
"kl_divergence": 5.827190534546389e-07,
"length_deviation": 0.0059127239367681595,
"vector_index": null,
"parameters": {
"attn.k_proj": {
"max_weight": 2.361252843342689,
"max_weight_position": 44.03583790301133,
"min_weight": 1.9866792183933073,
"min_weight_distance": 10.401301181948186
},
"attn.o_proj": {
"max_weight": 4.386315567926298,
"max_weight_position": 44.61946988151585,
"min_weight": 1.2438676170653118,
"min_weight_distance": 22.991611944992506
},
"attn.q_proj": {
"max_weight": 5.2791760970292945,
"max_weight_position": 42.172735552094444,
"min_weight": 0.08451546311113046,
"min_weight_distance": 11.801558054052062
},
"attn.v_proj": {
"max_weight": 1.04106066378018,
"max_weight_position": 50.31354129147488,
"min_weight": 0.22574977102298002,
"min_weight_distance": 11.86950016476815
}
}
}
],
"baseline_refusals": 99,
"eval_prompt_count": 100,
"completed_trials": 60,
"optimization_trials_configured": 60,
"judge": "google/gemini-3-flash-preview",
"generation": {
"min_new_tokens": 100,
"max_new_tokens": 150
},
"classic_safe_probe_file": "eval/top3_classic_safe_prompts_test.json"
}