File size: 8,191 Bytes
9e9c9ba | 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 | {
"architectures": [
"IKNN-Rl1-A1ForCausalLM"
],
"model_type": "iknn",
"model_name": "IKNN-Rl1-A1",
"organization": "deepRcurs Labs",
"repo_id": "deeprcurs/IKNN-Rl1-A1",
"file_name": "IKNN-Rl1-A1-150M.iknn",
"file_format": ".iknn",
"file_magic": "IKNN",
"file_arch": "IKNN-Rl1-A1",
"file_size_mb": 42,
"file_tensors": 86,
"hidden_size": 768,
"intermediate_size": 3072,
"num_hidden_layers": 12,
"num_attention_heads": 12,
"num_key_value_heads": 12,
"vocab_size": 32000,
"max_position_embeddings": 2048,
"rms_norm_eps": 1e-06,
"rope_theta": 10000.0,
"torch_dtype": "float32",
"transformers_version": "4.40.0",
"num_real_parameters": 150000000,
"num_active_parameters": 34500000,
"packed_parameter_count": 150000000,
"parameter_breakdown": {
"SatU1_1bit_87pct": {
"params": 130500000,
"tensors": 50,
"size_mb": 12.3,
"bits_per_param": 1.0,
"role": "throughput — XNOR popcount 16 Giga/s AVX2 32 Giga/s AVX-512"
},
"NoeSA_24_4_58bit_9pct": {
"params": 13500000,
"tensors": 24,
"size_mb": 20.2,
"bits_per_param": 4.58,
"states": 24,
"packing": "13 values per 60-bit = 4.615 bit/param",
"lut": "LUT576",
"distribution": "truncated Gaussian S1-S4 dual-zero",
"role": "critical logic/code/math — variance >94%"
},
"Ntarra_DnA_3_17bit_4pct": {
"params": 6000000,
"tensors": 12,
"size_mb": 11.2,
"bits_per_param": 3.17,
"states": 9,
"packing": "2x9 pack 5+8 bits = 77",
"direction": ["NEG", "ZERO", "POS"],
"phase": ["PHI0 shift0", "PHI1 shift2", "PHI2 shift4"],
"compute": "shift+sign no multiply — W_active=Sign(D)*(X<<phi)",
"role": "phase modulation — entropy gate"
}
},
"quantization_config": {
"method": "tri-tier-phase-quantization",
"tri_tier": {
"tier1_SatU1": {
"bits": 1.0,
"pct": 87,
"packing_factor_1bit": 32,
"compute": "XNOR + popcount — AVX2 16 Giga/s, AVX-512 VPOPCNTDQ 32 Giga/s",
"kernel": "satu1_avx2.cpp / satu1_avx512.cpp"
},
"tier2_NoeSA_24": {
"bits": 4.58,
"pct": 9,
"states": 24,
"packing_factor_4bit": 13,
"packing_detail": "13 values per 60-bit — 60/13=4.615 bit/param — base-24",
"lut_size": 576,
"scale": ["S1", "S2", "S3", "S4"],
"operator": ["-1a", "0a", "1a", "-1b", "0b", "1b"],
"kernel": "noesa24_avx2.cpp / noesa24_avx512.cpp"
},
"tier3_Ntarra_DnA": {
"bits": 3.17,
"pct": 4,
"states": 9,
"packing_factor": 2,
"packing_detail": "2x9 pack 5+8 bits = 77 — Direction 3 x Phase 3",
"shift": [0, 2, 4],
"sign": ["NEG -40", "ZERO 0", "POS 40"],
"kernel": "ntarra_avx2.cpp / ntarra_avx512.cpp"
}
},
"passive_experts_bits": 1.0,
"active_experts_bits": 4.58,
"packing_factor_1bit": 32,
"packing_factor_4bit": 13,
"packing_factor_3bit": 2,
"rht": {
"enabled": true,
"outlier_flattening": "10 -> 5.07",
"norm_preservation": "L2 11.3007 diff 9.5e-07",
"kernel": "rht_avx2.cpp / rht_avx512.cpp",
"method": "Randomized Hadamard Transform — H^T H = I — ||H x||_2 = ||x||_2"
},
"pg_kvc": {
"enabled": true,
"saving": "94% — 6.1MB -> 338KB — target -80%",
"full_pipeline": "96% — 720x1-bit 280x2-bit",
"gating": "2-bit + 1-bit phase-gated entropy"
},
"pep": {
"enabled": true,
"two_stage": "bigram cheap <0.5% + low-rank d_model->16->1",
"accuracy": "100%"
},
"adlp": {
"enabled": true,
"dual_worker": "fast 800 SatU1 + slow 200 NoeSA — 1e6 TPS"
}
},
"moe_config": {
"num_experts": 8,
"num_experts_per_tok": 2,
"router": "MoE Router with PEP Entropy Gate — Phase-Gated",
"router_method": "Gate(X) = Top-K(Softmax(Wr·X + br)) — tau_low 0.5 tau_high 1.5"
},
"runtime": {
"main": "iknn.cpp",
"kernels": 8,
"avx2": "Ryzen5 5650U Zen3 6C DDR4 38GB/s — target 28-42 TPS MTP / 50-68 non-MTP",
"avx512": "Xeon 2 vCPU AVX-512 — target 65-90 TPS MTP / 120-165 non-MTP",
"matmul_150m": "1000 tokens 176ms 5681 TPS Xeon 2 vCPU",
"memory_150m": "26MB — 34.5M active",
"memory_19_5b_target": "4.12GB — 19.5B parametric + 8B N-Gram non-parametric — avg 1.6 bits/param — design target"
},
"training": {
"method": "Clean Mining — anonymous frontier synthesis — pointer: CM-V2-20260903-##51pct",
"v1": {
"dataset": "6000 examples — 1000 each logic/reasoning/coding/research/math/science",
"train": 5000,
"val": 1000,
"steps": 100,
"batch": 4,
"loss": "7.08 -> 0.41",
"val_loss": "6.59 -> 0.50"
},
"v2": {
"dataset": "10k hard — 2000 logic hard, 2000 reasoning hard, 2000 coding hard, 2000 research hard, 1000 math hard, 1000 science hard",
"train": 8000,
"val": 2000,
"size": "7.6MB train + 1.9MB val",
"steps": 500,
"batch": 8,
"temp_annealing": "1.0 -> 0.01",
"n_states": "2-state {0,3} first 250 then 3-state {0,2,4}",
"loss": "7.01 -> 0.20",
"val_loss": "6.44 -> 0.20",
"improvement": "51% reduction from V1 0.41",
"grad_norm": "0.40-1.0 stable PASS",
"checkpoint_pt": "checkpoints/IKNN-Rl1-A1-150M-agentic-v2.pt 21MB — .cache excluded — no snapshot interference",
"checkpoint_safetensors": "model.safetensors 20MB + IKNN-Rl1-A1-150M.safetensors + checkpoints/IKNN-Rl1-A1-150M.safetensors — .cache excluded"
}
},
"datasets": {
"master_v2": "deeprcurs/IKNN-Rl1-Dataset-Agentic-V2 — 8000 train + 2000 val",
"master_all": "deeprcurs/IKNN-Rl1-Dataset",
"per_jenis": {
"Logic": "deeprcurs/IKNN-Rl1-Dataset-Logic — 1589 train + 411 val",
"Reasoning": "deeprcurs/IKNN-Rl1-Dataset-Reasoning — 1604 train + 396 val",
"Coding": "deeprcurs/IKNN-Rl1-Dataset-Coding — 1601 train + 399 val",
"Research": "deeprcurs/IKNN-Rl1-Dataset-Research — 1600 train + 400 val",
"Math": "deeprcurs/IKNN-Rl1-Dataset-Math — 791 train + 209 val",
"Science": "deeprcurs/IKNN-Rl1-Dataset-Science — 815 train + 185 val"
}
},
"snapshot": {
"limit": "128MB or 10000 files",
"excluded": [".cache", ".venv", "node_modules", "build", "dist", "out", "target", "__pycache__"],
"checkpoint_location": ".cache/ — excluded — no interference — snapshot stays 1.8M-4.5M tidy",
"zip_rule": "1 zip only in root — iknn-living.zip fixed name replaced each milestone — exception manual zip per owner request"
},
"protocols": {
"checkpoint_dataset_publish": "internal/protocol/CHECKPOINT_DATASET_PUBLISH_PROTOCOL.md — checkpoint .pt + .safetensors + dataset per jenis + (check)-## + no snapshot interference",
"clean_mining": "internal/protocol/CLEAN_MINING_PROTOCOL.md — Clean Mining anonymous — no distillation mention in public — pointer CM-V2-20260903-##51pct",
"comment_sanitization": "internal/protocol/COMMENT_SANITIZATION_PROTOCOL.md — technical only in public — internal only in /internal — pointer ## if forced",
"publish_all_in_one": "internal/protocol/PUBLISH_ALL_IN_ONE_PROTOCOL.md — one-click publish"
},
"pointers": {
"clean_mining": "CM-V2-20260903-##51pct",
"check_marker": "(check)-M6-v2-Loss0.20-##20MB-20260903-221711",
"manual_zip": "MANUAL-ZIP-##NO-IKNN-20260903"
},
"hf_links": {
"model": "https://huggingface.co/deeprcurs/IKNN-Rl1-A1",
"checkpoint_pt": "https://huggingface.co/deeprcurs/IKNN-Rl1-A1/blob/main/checkpoints/IKNN-Rl1-A1-150M-agentic-v2.pt",
"checkpoint_safetensors": "https://huggingface.co/deeprcurs/IKNN-Rl1-A1/blob/main/checkpoints/IKNN-Rl1-A1-150M.safetensors",
"model_safetensors": "https://huggingface.co/deeprcurs/IKNN-Rl1-A1/blob/main/model.safetensors",
"architecture_dark": "https://huggingface.co/deeprcurs/IKNN-Rl1-A1/blob/main/docs/IKNN-Rl1-A1-architecture-dark.png",
"dataset_master_v2": "https://huggingface.co/datasets/deeprcurs/IKNN-Rl1-Dataset-Agentic-V2"
}
}
|