Remove build artifacts and recipe details from comments and cards

JiRackDeltaNet_70b.py

@@ -1,8 +1,8 @@
 """
-JiRackDeltaNet_27b.py
+JiRackDeltaNet_70b.py
 =============================
 
-Собственная (hand-rolled) архитектура "JiRack Qwen3.8-27B" -- в стиле
+Собственная (hand-rolled) архитектура "JiRack" -- в стиле
 прежнего DS32B-проекта: один файл, свой TransformerBlock, никакого
 AutoConfig/AutoModelForCausalLM внутри класса архитектуры, никакого
 model_id в конструкторе. Все размерности зашиты константами ниже --
@@ -517,7 +517,6 @@ class TransformerBlock(nn.Module):
 
 class MTPBlock(nn.Module):
     """Multi-Token Prediction голова -- ровно как в реальном чекпоинте
-    Qwen3.5/3.8 (mtp_num_hidden_layers=1, mtp_use_dedicated_embeddings=false).
 
     Имена подмодулей СОВПАДАЮТ с ключами HF state_dict (top-level `mtp.`):
         mtp.pre_fc_norm_embedding.weight

NOTICE.md

@@ -3,7 +3,7 @@
 **JiRackDeltaNet 70B**
 Copyright 2026 Konstantin Vladimirovich Grabko (CMSManhattan)
 
-The model is distributed under the Business Source License 1.2. (see `LICENSE`).
+The model is distributed under the Business Source License 1.2 (see `LICENSE`).
 
 This model is a modified version of a Qwen model by the Qwen team, Alibaba Cloud, licensed under the Apache License,
 Version 2.0 (a copy is in `LICENSE-APACHE`). CMSManhattan modified the architecture and the weights and fine-tuned the

README.md

@@ -6,7 +6,6 @@ tags:
 - gated-deltanet
 - linear-attention
 - hybrid-attention
-- depth-expansion
 - function-calling
 - reasoning
 - fp8
@@ -129,7 +128,7 @@ This project is a derivative work based on a Qwen model originally released unde
 ### 1. Proprietary Terms for JiRack
 All modifications, proprietary additions, architectural extensions, derived model weights, and scripts developed by **CMS Manhattan (Konstantin Grabko)** — including but not limited to:
 * JiRack custom model code and architecture modifications,
-* FFN layer widening and depth expansion (up to 128 layers),
+* changes to the model architecture,
 * Supervised Fine-Tuning (SFT) on the JiRack dataset mix,
 * Quantized weights and conversions (FP8, GGUF, BitNet/ternary formats),
 
@@ -140,7 +139,7 @@ are licensed under the **JiRack Proprietary Model License (BSL)** (see `LICENSE`
 
 ### 2. Third-Party Attribution & Compliance
 In accordance with Section 4 of the Apache License 2.0, the original copyright, attribution, and license notices of the base model are preserved. When redistributing or using this repository, you must retain:
-* The original `LICENSE` (Apache 2.0) file,
+* The Apache 2.0 license text of the base model (`LICENSE-APACHE`),
 * The `NOTICE.md` file accompanying this repository,
 * All copyright and attribution notices attached to the original components.
 

audit_depth.json

@@ -1,14 +0,0 @@
-{
-  "kind": "depth",
-  "src": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded",
-  "out": "/mnt/nfs_share/JiRackDeltaNet_70b/expanded",
-  "tensors": 1714,
-  "ops": {
-    "copy": 1586,
-    "zero": 128
-  },
-  "errors": [],
-  "n_errors": 0,
-  "verdict": "PASS",
-  "minutes": 9.8
-}

chat_jirack.py

@@ -1,9 +1,8 @@
 #!/usr/bin/env python3
 # =============================================================================
-# chat_jirack.py - chat with JiRackDeltaNet (27B / 36B / 70B) from a safetensors folder.
+# chat_jirack.py - chat with JiRackDeltaNet 70B from a safetensors folder.
 #
 #   python3 chat_jirack.py --model-dir /mnt/nfs_share/JiRackDeltaNet_70b/expanded
-#   python3 chat_jirack.py --model-dir /mnt/nfs_share/JiRackDeltaNet_27b_cp      # 27B
 #
 # - model code: JiRackDeltaNet_*.py from the same folder (or --model-file)
 # - several GPUs: the TransformerBlocks are split across them evenly (70B bf16 = 140 GB,

gate_full.json

@@ -1,81 +0,0 @@
-{
-  "verdict": "PASS",
-  "expected": "exact",
-  "mini": false,
-  "compute": "fp32blocks",
-  "a": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded",
-  "b": "/mnt/nfs_share/JiRackDeltaNet_70b/expanded",
-  "layers_run": [
-    64,
-    128
-  ],
-  "intermediate": [
-    26624,
-    26624
-  ],
-  "prompts": [
-    "The capital of France is",
-    "def fibonacci(n):\n    ",
-    "Столица России —"
-  ],
-  "ids": [
-    [
-      760,
-      6511,
-      314,
-      9338,
-      369
-    ],
-    [
-      727,
-      73111,
-      1393,
-      1590,
-      198,
-      257
-    ],
-    [
-      212396,
-      168314,
-      170312,
-      1892
-    ]
-  ],
-  "rows": [
-    {
-      "T": 5,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 5
-    },
-    {
-      "T": 6,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 6
-    },
-    {
-      "T": 4,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 4
-    }
-  ],
-  "a_info": {
-    "cached": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded/gate_logits_fp32blocks_L64.npz"
-  },
-  "b_info": {
-    "loaded": 1714,
-    "skipped": 0,
-    "unexpected": [],
-    "missing_buffers": [],
-    "seconds": 2016
-  },
-  "torch": "2.13.0+cu130"
-}

gate_mini.json

@@ -1,81 +0,0 @@
-{
-  "verdict": "PASS",
-  "expected": "exact",
-  "mini": true,
-  "compute": "fp32blocks",
-  "a": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded",
-  "b": "/mnt/nfs_share/JiRackDeltaNet_70b/expanded",
-  "layers_run": [
-    4,
-    8
-  ],
-  "intermediate": [
-    26624,
-    26624
-  ],
-  "prompts": [
-    "The capital of France is",
-    "def fibonacci(n):\n    ",
-    "Столица России —"
-  ],
-  "ids": [
-    [
-      760,
-      6511,
-      314,
-      9338,
-      369
-    ],
-    [
-      727,
-      73111,
-      1393,
-      1590,
-      198,
-      257
-    ],
-    [
-      212396,
-      168314,
-      170312,
-      1892
-    ]
-  ],
-  "rows": [
-    {
-      "T": 5,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 5
-    },
-    {
-      "T": 6,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 6
-    },
-    {
-      "T": 4,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 4
-    }
-  ],
-  "a_info": {
-    "cached": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded/gate_logits_fp32blocks_L4.npz"
-  },
-  "b_info": {
-    "loaded": 124,
-    "skipped": 1590,
-    "unexpected": [],
-    "missing_buffers": [],
-    "seconds": 138
-  },
-  "torch": "2.13.0+cu130"
-}

gguf.txt

@@ -1,35 +0,0 @@
-# 1. Конвертация в bf16 (как договорились)
-python convert_hf_to_gguf.py /mnt/nfs_share/Qwen3_8 \
-    --outfile /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.gguf --outtype bf16
-
-# 2. Сборка llama-quantize (один раз)
-cd /mnt/nfs_share/llama.cpp
-cmake -B build
-cmake --build build -j
-
-# 3. Квантование в Q4_K_M
-./build/bin/llama-quantize \
-    /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.gguf \
-    /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.Q4_K_M.gguf \
-    Q4_K_M
-
-
-
-cd /mnt/nfs_share/llama.cpp
-cmake -B build -DGGML_CUDA=ON  # для NVIDIA
-make -C build
-
-
-/mnt/nfs_share/llama.cpp/build/bin/llama-cli \
-     -m /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.Q4_K_M.gguf \
-     -n 512 \
-     -c 4096 \
-     -p "Explain quantum computing in simple terms:"
-
-
-
-
-
-/mnt/nfs_share/llama.cpp/build/bin/llama-server \
-     -m /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.Q4_K_M.gguf \
-     --port 8000

jirack_sft_state.json

@@ -1,9 +0,0 @@
-{
-  "global_step": 374,
-  "lr": 1e-05,
-  "max_steps": 374,
-  "shards": "0",
-  "tok_micro": 8192,
-  "accum": 2,
-  "init_dir": "/home/ubuntu/deltanet70b"
-}

sft_final_fp8/JiRackDeltaNet_70b.py

@@ -1,8 +1,8 @@
 """
-JiRackDeltaNet_27b.py
+JiRackDeltaNet_70b.py
 =============================
 
-Собственная (hand-rolled) архитектура "JiRack Qwen3.8-27B" -- в стиле
+Собственная (hand-rolled) архитектура "JiRack" -- в стиле
 прежнего DS32B-проекта: один файл, свой TransformerBlock, никакого
 AutoConfig/AutoModelForCausalLM внутри класса архитектуры, никакого
 model_id в конструкторе. Все размерности зашиты константами ниже --
@@ -517,7 +517,6 @@ class TransformerBlock(nn.Module):
 
 class MTPBlock(nn.Module):
     """Multi-Token Prediction голова -- ровно как в реальном чекпоинте
-    Qwen3.5/3.8 (mtp_num_hidden_layers=1, mtp_use_dedicated_embeddings=false).
 
     Имена подмодулей СОВПАДАЮТ с ключами HF state_dict (top-level `mtp.`):
         mtp.pre_fc_norm_embedding.weight

sft_final_fp8/audit_depth.json

@@ -1,14 +0,0 @@
-{
-  "kind": "depth",
-  "src": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded",
-  "out": "/mnt/nfs_share/JiRackDeltaNet_70b/expanded",
-  "tensors": 1714,
-  "ops": {
-    "copy": 1586,
-    "zero": 128
-  },
-  "errors": [],
-  "n_errors": 0,
-  "verdict": "PASS",
-  "minutes": 9.8
-}

sft_final_fp8/chat_jirack.py

@@ -1,9 +1,8 @@
 #!/usr/bin/env python3
 # =============================================================================
-# chat_jirack.py - chat with JiRackDeltaNet (27B / 36B / 70B) from a safetensors folder.
+# chat_jirack.py - chat with JiRackDeltaNet 70B from a safetensors folder.
 #
 #   python3 chat_jirack.py --model-dir /mnt/nfs_share/JiRackDeltaNet_70b/expanded
-#   python3 chat_jirack.py --model-dir /mnt/nfs_share/JiRackDeltaNet_27b_cp      # 27B
 #
 # - model code: JiRackDeltaNet_*.py from the same folder (or --model-file)
 # - several GPUs: the TransformerBlocks are split across them evenly (70B bf16 = 140 GB,

sft_final_fp8/gate_full.json

@@ -1,81 +0,0 @@
-{
-  "verdict": "PASS",
-  "expected": "exact",
-  "mini": false,
-  "compute": "fp32blocks",
-  "a": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded",
-  "b": "/mnt/nfs_share/JiRackDeltaNet_70b/expanded",
-  "layers_run": [
-    64,
-    128
-  ],
-  "intermediate": [
-    26624,
-    26624
-  ],
-  "prompts": [
-    "The capital of France is",
-    "def fibonacci(n):\n    ",
-    "Столица России —"
-  ],
-  "ids": [
-    [
-      760,
-      6511,
-      314,
-      9338,
-      369
-    ],
-    [
-      727,
-      73111,
-      1393,
-      1590,
-      198,
-      257
-    ],
-    [
-      212396,
-      168314,
-      170312,
-      1892
-    ]
-  ],
-  "rows": [
-    {
-      "T": 5,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 5
-    },
-    {
-      "T": 6,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 6
-    },
-    {
-      "T": 4,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 4
-    }
-  ],
-  "a_info": {
-    "cached": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded/gate_logits_fp32blocks_L64.npz"
-  },
-  "b_info": {
-    "loaded": 1714,
-    "skipped": 0,
-    "unexpected": [],
-    "missing_buffers": [],
-    "seconds": 2016
-  },
-  "torch": "2.13.0+cu130"
-}

sft_final_fp8/gate_mini.json

@@ -1,81 +0,0 @@
-{
-  "verdict": "PASS",
-  "expected": "exact",
-  "mini": true,
-  "compute": "fp32blocks",
-  "a": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded",
-  "b": "/mnt/nfs_share/JiRackDeltaNet_70b/expanded",
-  "layers_run": [
-    4,
-    8
-  ],
-  "intermediate": [
-    26624,
-    26624
-  ],
-  "prompts": [
-    "The capital of France is",
-    "def fibonacci(n):\n    ",
-    "Столица России —"
-  ],
-  "ids": [
-    [
-      760,
-      6511,
-      314,
-      9338,
-      369
-    ],
-    [
-      727,
-      73111,
-      1393,
-      1590,
-      198,
-      257
-    ],
-    [
-      212396,
-      168314,
-      170312,
-      1892
-    ]
-  ],
-  "rows": [
-    {
-      "T": 5,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 5
-    },
-    {
-      "T": 6,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 6
-    },
-    {
-      "T": 4,
-      "identical": true,
-      "max_abs": 0.0,
-      "mean_abs": 0.0,
-      "max_kl": 0.0,
-      "top1_agree": 4
-    }
-  ],
-  "a_info": {
-    "cached": "/mnt/nfs_share/JiRackDeltaNet_36b/expanded/gate_logits_fp32blocks_L4.npz"
-  },
-  "b_info": {
-    "loaded": 124,
-    "skipped": 1590,
-    "unexpected": [],
-    "missing_buffers": [],
-    "seconds": 138
-  },
-  "torch": "2.13.0+cu130"
-}

sft_final_fp8/gguf.txt

@@ -1,35 +0,0 @@
-# 1. Конвертация в bf16 (как договорились)
-python convert_hf_to_gguf.py /mnt/nfs_share/Qwen3_8 \
-    --outfile /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.gguf --outtype bf16
-
-# 2. Сборка llama-quantize (один раз)
-cd /mnt/nfs_share/llama.cpp
-cmake -B build
-cmake --build build -j
-
-# 3. Квантование в Q4_K_M
-./build/bin/llama-quantize \
-    /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.gguf \
-    /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.Q4_K_M.gguf \
-    Q4_K_M
-
-
-
-cd /mnt/nfs_share/llama.cpp
-cmake -B build -DGGML_CUDA=ON  # для NVIDIA
-make -C build
-
-
-/mnt/nfs_share/llama.cpp/build/bin/llama-cli \
-     -m /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.Q4_K_M.gguf \
-     -n 512 \
-     -c 4096 \
-     -p "Explain quantum computing in simple terms:"
-
-
-
-
-
-/mnt/nfs_share/llama.cpp/build/bin/llama-server \
-     -m /mnt/nfs_share/Qwen3_8/JiRackDeltaNet_27b.Q4_K_M.gguf \
-     --port 8000

sft_final_fp8/jirack_sft_state.json

@@ -1,9 +0,0 @@
-{
-  "global_step": 374,
-  "lr": 1e-05,
-  "max_steps": 374,
-  "shards": "0",
-  "tok_micro": 8192,
-  "accum": 2,
-  "init_dir": "/home/ubuntu/deltanet70b"
-}