| { |
| "model": "allenai/OLMo-2-0425-1B-Instruct", |
| "study_checkpoint_dir": "/common/users/vp752/iconoclast_ilabs/checkpoints/olmo2-1b-seq", |
| "base_metrics": { |
| "refusals": 17, |
| "overrefusals": 1, |
| "harmful_marker_hits": 49, |
| "harmful_compliance_score": 0.5791666666666666, |
| "objective_regime": "refusal_reduction" |
| }, |
| "pareto_trials": [ |
| { |
| "index": 43, |
| "refusals": 2, |
| "overrefusals": 0, |
| "harmful_marker_hits": 2, |
| "harmful_compliance_score": 0.8798958333333331, |
| "objective_regime": "refusal_reduction", |
| "merge_penalty": 0.0, |
| "kl_divergence": 0.03453812748193741, |
| "direction_method": "hybrid", |
| "direction_scope": "global", |
| "direction_index": 6.803772164857431, |
| "direction_blend": 0.8126204340417661, |
| "parameters": { |
| "attn.o_proj": { |
| "max_weight": 1.9846663609335586, |
| "max_weight_position": 8.470516856790118, |
| "min_weight": 0.1346933022408364, |
| "min_weight_distance": 5.814092682057282 |
| }, |
| "mlp.down_proj": { |
| "max_weight": 1.3370699613454922, |
| "max_weight_position": 7.388872181956926, |
| "min_weight": 1.015835112416409, |
| "min_weight_distance": 8.500454138090097 |
| } |
| }, |
| "harmful_axis_metrics": {} |
| }, |
| { |
| "index": 8, |
| "refusals": 5, |
| "overrefusals": 0, |
| "harmful_marker_hits": 8, |
| "harmful_compliance_score": 0.8578124999999999, |
| "objective_regime": "refusal_reduction", |
| "merge_penalty": 0.0, |
| "kl_divergence": 0.022717345505952835, |
| "direction_method": "mean", |
| "direction_scope": "global", |
| "direction_index": 6.888988867439076, |
| "direction_blend": 0.7501162670210468, |
| "parameters": { |
| "attn.o_proj": { |
| "max_weight": 1.5027759251817483, |
| "max_weight_position": 11.439685163347425, |
| "min_weight": 0.1404937683129481, |
| "min_weight_distance": 8.848109589720098 |
| }, |
| "mlp.down_proj": { |
| "max_weight": 1.8833228628835126, |
| "max_weight_position": 8.723217139252817, |
| "min_weight": 0.32624355530269145, |
| "min_weight_distance": 6.482577936343204 |
| } |
| }, |
| "harmful_axis_metrics": {} |
| } |
| ] |
| } |
