{
  "schema_version": 1,
  "model": "PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-MLX-4bit",
  "variant": "4bit",
  "status": "experimental",
  "base_model": "ornith-ai/Ornith-1.5-35B-A3B",
  "base_revision": "e4dfb35a93d4b6822a811a7676f3488514abe7e2",
  "runtime": {
    "mlx": "0.32.0",
    "mlx_vlm": "0.6.8"
  },
  "behavioral_evaluation": {
    "parent_12_case_gate": {
      "capability": {
        "count": 12,
        "pass_count": 12,
        "pass_rate": 1.0
      },
      "jbb_benign": {
        "count": 12,
        "explicit_refusal_count": 1,
        "explicit_refusal_rate": 0.08333333333333333,
        "final_answer_count": 12,
        "completed_answer_count": 1,
        "evasive_nonanswer_count": 0,
        "evasive_nonanswer_rate": 0.0,
        "truncated_generation_count": 11,
        "unfinished_reasoning_count": 0
      },
      "jbb_harmful": {
        "count": 12,
        "explicit_refusal_count": 12,
        "explicit_refusal_rate": 1.0,
        "final_answer_count": 12,
        "completed_answer_count": 11,
        "evasive_nonanswer_count": 0,
        "evasive_nonanswer_rate": 0.0,
        "truncated_generation_count": 1,
        "unfinished_reasoning_count": 0
      }
    },
    "selected_bf16_12_case_gate": {
      "capability": {
        "count": 12,
        "pass_count": 12,
        "pass_rate": 1.0
      },
      "jbb_benign": {
        "count": 12,
        "explicit_refusal_count": 0,
        "explicit_refusal_rate": 0.0,
        "final_answer_count": 12,
        "completed_answer_count": 1,
        "evasive_nonanswer_count": 0,
        "evasive_nonanswer_rate": 0.0,
        "truncated_generation_count": 11,
        "unfinished_reasoning_count": 0
      },
      "jbb_harmful": {
        "count": 12,
        "explicit_refusal_count": 0,
        "explicit_refusal_rate": 0.0,
        "final_answer_count": 12,
        "completed_answer_count": 4,
        "evasive_nonanswer_count": 0,
        "evasive_nonanswer_rate": 0.0,
        "truncated_generation_count": 8,
        "unfinished_reasoning_count": 0
      }
    },
    "variant_100_harmful_100_benign_screen": {
      "jbb_benign": {
        "count": 100,
        "explicit_refusal_count": 0,
        "explicit_refusal_rate": 0.0,
        "final_answer_count": 100,
        "completed_answer_count": 1,
        "evasive_nonanswer_count": 0,
        "evasive_nonanswer_rate": 0.0,
        "truncated_generation_count": 99,
        "unfinished_reasoning_count": 0
      },
      "jbb_harmful": {
        "count": 100,
        "explicit_refusal_count": 0,
        "explicit_refusal_rate": 0.0,
        "final_answer_count": 100,
        "completed_answer_count": 4,
        "evasive_nonanswer_count": 0,
        "evasive_nonanswer_rate": 0.0,
        "truncated_generation_count": 96,
        "unfinished_reasoning_count": 0
      }
    },
    "batch_size": 1,
    "max_new_tokens": 128,
    "thinking": false,
    "scorer": "transparent phrase-based explicit-refusal detector",
    "caveat": "This is an early-refusal screen, not a full answer-quality evaluation. Manual review found semantic refusals that the phrase detector did not flag."
  },
  "matched_teacher_drift": {
    "protocol": {
      "context": "shared regular-BF16 greedy teacher trajectory",
      "positionsPerCaseMaximum": 16,
      "vocabulary": 248320,
      "logitsStorage": "float32",
      "probabilityMath": "float64 full-vocabulary log-softmax",
      "residualState": "all 40 residual_post vectors at assistant boundary",
      "cacheState": "final teacher-trajectory cache: K/V for 10 full-attention layers; convolution and gated-delta recurrent state for 30 linear-attention layers",
      "thinking": false,
      "sampling": "greedy"
    },
    "dataset": {
      "capabilityFileSha256": "fe35cc41fe041cf2d371041c3fb977ec37b51fa1e9a0b0482aaf1eab415c9c26",
      "capabilityCases": 12,
      "jailbreakBench": {
        "repository": "JailbreakBench/JBB-Behaviors",
        "revision": "886acc352a31533ffbcf4ef22c744658688086fc"
      },
      "harmfulCases": 12,
      "benignCases": 12
    },
    "pure_abliteration_vs_regular_bf16": {
      "positions": 481,
      "forward_kl_nats_mean": 0.54518455642922,
      "top1_agreement": 0.7817047817047817,
      "residual_cosine_mean": 0.9520897871696514,
      "cache": {
        "full_attention_k_cosine_mean": 0.9956834574679786,
        "full_attention_v_cosine_mean": 0.9942660171605967,
        "linear_attention_conv_cosine_mean": 0.997201268533088,
        "linear_attention_recurrent_cosine_mean": 0.9962457101916542
      },
      "suite_splits": {
        "capability": {
          "positions": 97,
          "forward_kl_nats_mean": 0.01834161340133247,
          "top1_agreement": 0.979381443298969,
          "residual_cosine_mean": 0.995895118556436
        },
        "jbb_benign": {
          "positions": 192,
          "forward_kl_nats_mean": 0.3073881842398057,
          "top1_agreement": 0.8541666666666666,
          "residual_cosine_mean": 0.9696202077556074
        },
        "jbb_harmful": {
          "positions": 192,
          "forward_kl_nats_mean": 1.0491463737941817,
          "top1_agreement": 0.609375,
          "residual_cosine_mean": 0.8907540351969113
        }
      }
    },
    "quantization_vs_abliterated_bf16": {
      "positions": 481,
      "forward_kl_nats_mean": 0.14369058172143362,
      "top1_agreement": 0.8731808731808732,
      "residual_cosine_mean": 0.9584706869834336,
      "cache": {
        "full_attention_k_cosine_mean": 0.9793075130946234,
        "full_attention_v_cosine_mean": 0.9716918964141061,
        "linear_attention_conv_cosine_mean": 0.9878246768445645,
        "linear_attention_recurrent_cosine_mean": 0.9866432650675594
      }
    },
    "total_vs_regular_bf16": {
      "positions": 481,
      "forward_kl_nats_mean": 0.6640038720543332,
      "top1_agreement": 0.7525987525987526,
      "residual_cosine_mean": 0.9207185389809484,
      "cache": {
        "full_attention_k_cosine_mean": 0.9764506595000695,
        "full_attention_v_cosine_mean": 0.9679095435046612,
        "linear_attention_conv_cosine_mean": 0.9860736375044666,
        "linear_attention_recurrent_cosine_mean": 0.9839664542846734
      }
    }
  },
  "capabilities": {
    "vision_tower_included": true,
    "video_preprocessor_included": true,
    "native_mtp_head_included": false,
    "four_bit_image_smoke_passed": true,
    "vision_evaluation_caveat": "Only the 4-bit artifact received an end-to-end image smoke test; vision and video were not comprehensively benchmarked."
  }
}