LoganResearch/Adaptive-Repetition-Controller-ARC

1{2  "model_type": "adaptive_repetition_controller",3  "version": "1.0.0",4  "architecture": {5    "d_model": 4096,6    "n_layers": 32,7    "d_fiber": 16,8    "d_control": 64,9    "rep_window": 32,10    "total_params": 5000011  },12  "training": {13    "dataset": "wikitext-2",14    "loss": "BCEWithLogitsLoss",15    "pos_weight": "dynamic",16    "lr_predictor": 1e-4,17    "lr_lora": 2e-5,18    "batch_size": 4,19    "gradient_accumulation": 8,20    "optimal_steps": 500021  },22  "performance": {23    "f1_score": 0.99,24    "risk_at_repeats": 0.998,25    "risk_at_non_repeats": 0.008,26    "separation": "125x",27    "repetition_reduction": "48.4%",28    "distinct2_improvement": "16.7%"29  },30  "inference": {31    "penalty_scale_default": 3.0,32    "temperature_default": 0.8,33    "threshold_default": 0.1,34    "rep_window": 3235  },36  "base_model_compatibility": [37    "llama-3.1-8b",38    "llama-3-8b",39    "mistral-7b"40  ],41  "notes": "This is a decode-time intervention system, not an attention modification. The geometric CF-HoT theory remains unvalidated; this is the working practical implementation."42}43