spb/modelmap Public License

1{2  "map_id": "atlas/qwen3-0.6b-4bit/probes/v1",3  "map_type": "probes",4  "model_id": "mlx-community/Qwen3-0.6B-4bit",5  "model_hash": "392e8d466d56100ada00eb82031fb854297fc9e389b7d303eba3af114e87bce2",6  "quantization": "q4 (mlx)",7  "commit": "3935e7933294b9c1293cd31b887126402fc53115",8  "config": "results/expA_probe_reliability/20260812T062605Z/results.json",9  "created": "2026-08-12",10  "hardware_manifest": {11    "author": "Simon-Pierre Boucher",12    "contact": "contact@spboucher.ai",13    "website": "https://modelmap.io",14    "chip": {15      "brand": "Apple M5 Max",16      "cores_total": 18,17      "cores_performance": 6,18      "cores_efficiency": 1219    },20    "memory": {21      "unified_gb": 48.0,22      "pagesize": 1638423    },24    "os": {25      "system": "Darwin",26      "version": "27.0",27      "arch": "arm64"28    },29    "software": {30      "python": "3.14.4",31      "numpy": "2.5.2",32      "mlx": "0.32.0",33      "torch": "2.13.0",34      "safetensors": "0.8.0"35    }36  },37  "confidence_level": 1,38  "regenerate_command": ".venv/bin/python benchmarks/promptsets/make_promptsets.py && .venv/bin/python experiments/micro/expA_probe_reliability/implementation/benchmark.py && .venv/bin/python experiments/micro/expA_probe_reliability/implementation/make_mapcard.py",39  "seeds": [40    0,41    1,42    2,43    3,44    445  ],46  "prompt_sets": [47    "arith_A.jsonl#2fd80600d8a1b4ad",48    "arith_B.jsonl#04bb0eceb4b9262e",49    "code_prose_A.jsonl#dfbfb13dade0fd0b",50    "code_prose_B.jsonl#e9c3f78d8754b0ad",51    "lang_id_A.jsonl#43bd7ed12d2d9f95",52    "lang_id_B.jsonl#833435dae9a61ae0"53  ],54  "controls": [55    "shuffled-label (every probe)",56    "random-init architecture twin",57    "BH-FDR q=0.05 across layer scans"58  ],59  "methods_in_agreement": [],60  "interventions": [],61  "replication_rate": 1.0,62  "per_dataset_agreement": null,63  "ablation_schemes": [],64  "featurizer_class": "natural-basis (mean-pooled residual)",65  "intervention_protocol": "none (observational map — Level 1 by design)",66  "negative_result": true,67  "notes": "NEGATIVE RESULT: the random-init architecture twin reaches task accuracy 1.00 at every layer for every property — the probe map is indistinguishable from the architecture+tokenizer null on these template promptsets. This map is published as evidence that probe maps on lexically separable classes are uninformative about trained structure. See analysis.md.",68  "author": "Simon-Pierre Boucher",69  "contact": "contact@spboucher.ai",70  "website": "https://modelmap.io",71  "schema_version": "0.1"72}73