File size: 6,087 Bytes
f1ce0ae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 | {
"code_termination_diagnostic": {
"among_outputs_reaching_execution": {
"control": {
"executed": 103,
"pass_percent": 15.53,
"passed": 16
},
"winner": {
"executed": 46,
"pass_percent": 21.74,
"passed": 10
}
},
"interpretation": "The full-code result is dominated by severe termination/extraction pathology and is not a clean latent-code estimate.",
"winner_generations_hitting_512_token_cap": 376,
"winner_generations_total": 421
},
"comparison": {
"benign_kl": {
"control": 0.0,
"strict_limit": 0.05,
"winner": 0.093614
},
"capability_macro_percent": {
"control": 17.6859,
"winner": 21.0086
},
"exact_prompt_echoes": {
"control": 2,
"leakage_detected": true,
"winner": 3
},
"full_code_passes": {
"control": 16,
"denominator": 421,
"winner": 10
},
"harmful_hard_refusal_percent": {
"control": 43.2,
"winner": 0.0
},
"harmful_soft_deflection_percent": {
"control": 14.6,
"winner": 0.2
},
"harmful_substantive_response_percent": {
"control": 47.0,
"winner": 99.4
},
"held_out_loss_ratio": {
"control": 1.0,
"limit": 1.05,
"winner": 1.024478
},
"human_eval_percent": {
"control": 7.9268,
"winner": 4.2683
},
"incoherence_percent": {
"control": 2.7692,
"strict_winner_limit": 2.7692,
"winner": 4.3077
},
"long_form_pass_percent": {
"control": 54.1667,
"winner": 62.5
},
"max_repeated_4gram_fraction_percent": {
"strict_limit": 5.0,
"winner": 5.8632
},
"mmmu30_correct": {
"control": 9,
"denominator": 30,
"winner": 11
}
},
"evaluation_provenance": {
"kind": "self-run_frozen_local_benchmarks",
"note": "Results were produced by this project on a frozen local evaluation suite. They must not be mixed with or represented as official Qwen benchmark results.",
"official_qwen_benchmarks": false,
"release_report_frozen_at": "2026-08-22T00:01:48.742409+00:00"
},
"lineage": {
"abliterix_version": "1.12.2",
"base_model": "Qwen/Qwen3.8-27B",
"method": "Reasoning QLoRA merge followed by one selected Abliterix pass for the winner; control received no Abliterix residual-writer edits.",
"reasoning_control": "control-bf16",
"seed": 42,
"winner": "abliterix-pass1-bf16"
},
"native_mlx_proof": {
"control_mtp_statistics": {
"accepted_drafts_per_round": 1.73,
"accepted_tokens_per_round": 2.73,
"average_draft": 2.0,
"drafted_percent": 86.5,
"rounds": 37
},
"mtp_exact_answer": {
"control": "323",
"winner": "323"
},
"ordinary_exact_answer": {
"control": "323",
"winner": "323"
},
"runtime": "mlx-vlm on arm64 macOS/Metal",
"speed_note": "MTP speed is workload-dependent; positive acceptance and exact-answer agreement are the correctness evidence.",
"winner_mtp_statistics": {
"accepted_drafts_per_round": 1.78,
"accepted_tokens_per_round": 2.78,
"average_draft": 2.0,
"drafted_percent": 88.8,
"rounds": 76
},
"winner_quantization": {
"bits": 8,
"group_size": 64,
"scheme": "affine"
}
},
"release_family": "Qwen3.8-27B Opus Personal Model v1",
"schema_version": 1,
"selection": {
"all_preregistered_strict_gates_passed": false,
"control_variant": "control-bf16",
"selected_variant": "abliterix-pass1-bf16",
"status": "selected_practical_winner_with_measured_deviations",
"summary": "Practical personal-model selection with measured deviations; not universal dominance."
},
"strict_deviations": [
{
"metric": "benign_kl",
"passed": false,
"strict_limit": 0.05,
"winner": 0.093614
},
{
"metric": "incoherence_percent",
"passed": false,
"strict_limit": 2.7692,
"winner": 4.3077
},
{
"control": 7.9268,
"metric": "human_eval_percent",
"passed": false,
"strict_delta_floor": -3.0,
"winner": 4.2683,
"winner_delta_points": -3.6585
},
{
"control": "16/421",
"metric": "full_code_passes",
"passed": false,
"winner": "10/421"
},
{
"metric": "max_repeated_4gram_fraction_percent",
"passed": false,
"strict_limit": 5.0,
"winner": 5.8632
},
{
"control_exact_echoes": 2,
"metric": "prompt_leakage",
"passed": false,
"winner_exact_echoes": 3
}
],
"tensor_integrity": {
"native_mtp_tensors": 15,
"total_tensor_keys": 1199,
"unexpected_changes": 0,
"vision_tensors_preserved": 333,
"winner_abliterix_residual_writer_edits": 74
},
"training": {
"dataset_preparation": {
"accepted_rows": 12614,
"duplicates_removed": 208,
"invalid_rows_removed": 20,
"processed_manifest_sha256": "6e0a36ad20732c5f98ff592c4565a4c86876fead4e9f94bc6ceedfad1339a94d",
"raw_rows": 12842,
"rows_published": false,
"source_rows": {
"high-reasoning-250x": 250,
"opus-10000x": 9633,
"opus-3000x": 2326,
"reasoning-700x": 633
},
"split_sha256": {
"test.jsonl": "a2b6b46ce8b054166b0e392701e49390a44e391b56d25a0a25e9b8a9000257d9",
"train.jsonl": "c7e691bcabd945e1259537f74119bce4c2c8f5fa4d2fba2e626b95bc04176ad3",
"validation.jsonl": "589a05d0cf2fededff1febb065d6b3c8c947d29e948495c1567bcd9f5c44f364"
},
"splits": {
"test": 138,
"train": 12349,
"validation": 127
}
},
"dtype_after_merge": "bfloat16",
"final_validation_loss": 0.23739749,
"hardware": "NVIDIA H200-class run; the sealed Abliterix environment recorded NVIDIA H200, CUDA 13.0, driver 580.159.03, Python 3.11.15.",
"optimizer_steps": 1544,
"rows": 12349,
"token_accuracy_percent": 91.7594,
"trainable_lora_parameters": 108789760
}
}
|