builds/nemotron-omni-moe: nemotron_h_moe MoE shares on the model's own markup

Recipe nemotron-omni-moe (experiments/calib_nemotron/recipes/nemotron-moe.yaml in atomic-quantizer), rendered through the model's own chat_template.jinja (auto_fmt), its vocabulary sweep (131072 rows, the Lightning tokenizer). Pool documents sharing a 13-gram with an eval file were left out of the build (purge_overlap); the pool itself is unchanged by this commit.

builds/nemotron-omni-moe/calib_longctx.manifest.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:354542e7b40faa618106cc1bdb91597d158ca74464f6a86d0f95ac59b3f93292
+size 7249

builds/nemotron-omni-moe/calib_longctx.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f04ebaf4030b3b5e4d785b4a7d326a907a053de7c4567f8d809bceca9a959007
+size 2798988

builds/nemotron-omni-moe/calib_train.manifest.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:299552ae568484f1b2ae0da2d9bbeb617ce4f66a744d53bb2a266e2d7acae381
+size 1019098

builds/nemotron-omni-moe/calib_train.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c4fd31a4096783d2a2bd0e7bae53b4108efab61436e4389eef567834e2be65e
+size 16648839

builds/nemotron-omni-moe/manifest.json

@@ -0,0 +1,277 @@
+{
+  "build": "nemotron-omni-moe",
+  "built_at": null,
+  "build_tool_version": "build.py/1.0",
+  "recipe": {
+    "path": "recipes/nemotron-omni-moe.yaml",
+    "sha256": "c2dff0a3e1844d5b6fdacfbe2c7aa58a915c9c0660b5c7f18bf673fc22e15557",
+    "seed": 20261015
+  },
+  "model": {
+    "repo": "nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16",
+    "tokenizer": "/src/tokenizer.json",
+    "vocab_size": 131072,
+    "layers": 88,
+    "sliding_window": null,
+    "context_length": 262144,
+    "llama_cpp_commit": "e64207bab6c2"
+  },
+  "tokenizer": {
+    "path": "/src/tokenizer.json",
+    "sha256": "d0432ad8b1c2984a2e2dca82e1933e3fd5628df6e95c51fba4c3ee9950dce299",
+    "rows": 131072,
+    "pre_tokenizer": "llama4"
+  },
+  "llama_cpp": {
+    "commit": "e64207bab6c2",
+    "note": "llama-imatrix must be given --parse-special or the chat markup in calib_train.txt is tokenised as literal text (parse_special defaults to false in that tool)"
+  },
+  "calib_train": {
+    "path": "calib_train.txt",
+    "sha256": "9c4fd31a4096783d2a2bd0e7bae53b4108efab61436e4389eef567834e2be65e",
+    "bytes": 16648839,
+    "documents": 2820,
+    "tokens": 4589438,
+    "target_tokens": 5000000,
+    "document_separator": "\n\n",
+    "shares_requested_percent": {
+      "agentic": 22.0,
+      "code": 16.0,
+      "graphics": 4.0,
+      "longctx": 10.0,
+      "multilingual": 18.0,
+      "reasoning": 12.0,
+      "structured": 6.0,
+      "vocab_sweep": 12.0
+    },
+    "shares_actual_percent": {
+      "agentic": 24.0,
+      "code": 17.441,
+      "graphics": 4.37,
+      "longctx": 11.145,
+      "multilingual": 19.639,
+      "reasoning": 13.091,
+      "structured": 4.393,
+      "vocab_sweep": 5.92
+    },
+    "documents_by_category": {
+      "agentic": 374,
+      "code": 501,
+      "graphics": 127,
+      "longctx": 22,
+      "multilingual": 214,
+      "reasoning": 1018,
+      "structured": 451,
+      "vocab_sweep": 113
+    },
+    "tokens_by_category": {
+      "agentic": 1101485,
+      "code": 800459,
+      "graphics": 200568,
+      "longctx": 511472,
+      "multilingual": 901311,
+      "reasoning": 600803,
+      "structured": 201626,
+      "vocab_sweep": 271714
+    },
+    "synthetic_tokens": 2035924,
+    "synthetic_percent": 44.361,
+    "document_tokens": {
+      "p50": 643,
+      "p90": 3908,
+      "p95": 6020,
+      "p99": 15868,
+      "max": 51537,
+      "min": 65
+    }
+  },
+  "pool_sources": {
+    "pool/_quarantine/eval-residue.jsonl": {
+      "sha256": "9e7703a3c205b425f83f17110cf7a00b7558f8ed501b45f1fcb0f4145f5f2579",
+      "bytes": 1635378,
+      "documents": 64
+    },
+    "pool/_quarantine/removed.jsonl": {
+      "sha256": "28615d8080e4b3b95a29447402e8163786c42e23e6c5175a18783a591bde13d0",
+      "bytes": 8600246,
+      "documents": 1056
+    },
+    "pool/_quarantine/untokenizable.jsonl": {
+      "sha256": "8226d1b281ce13ceab909c06383d78176b3cc5c2d3f1a49094ef808e3f220335",
+      "bytes": 88848,
+      "documents": 2
+    },
+    "pool/agentic/synthetic/conversations.jsonl": {
+      "sha256": "40478e56fdee9bf965b0e56800e3d940e772db11e856f0170d92043d86861973",
+      "bytes": 5641610,
+      "documents": 531
+    },
+    "pool/agentic/synthetic/legacy-dsv4-rendered.jsonl": {
+      "sha256": "24e99d36930e6547e90058654b712ae50a1294a1cdc4da7b38ae688a97537dca",
+      "bytes": 948845,
+      "documents": 50
+    },
+    "pool/code/legacy-dsv4.jsonl": {
+      "sha256": "28076d09ec882624d77b30289094ab65ae8c1fb57d3dceb55011b5e81fa99916",
+      "bytes": 471958,
+      "documents": 119
+    },
+    "pool/code/repos.jsonl": {
+      "sha256": "549d67465b052d78a4277da12033840309d767cd897f7d1280df408b43bbc916",
+      "bytes": 9133710,
+      "documents": 831
+    },
+    "pool/german/agentic_de.jsonl": {
+      "sha256": "f7afba5c9872c9e69f6fc9b0e4eb72dc1e80d8eaef1c92c736636b924eb00609",
+      "bytes": 1230904,
+      "documents": 735
+    },
+    "pool/german/germanrag.jsonl": {
+      "sha256": "23a5c7e8d2cea57ab8e01d5c8b953b81e6eb7811f12dae4b59b8396ceebf27ba",
+      "bytes": 1029498,
+      "documents": 265
+    },
+    "pool/german/longctx.jsonl": {
+      "sha256": "26d9b0f152105337fad88e098dcffa7dbcae8fe3eccd8c0a35850603b485ea60",
+      "bytes": 2488298,
+      "documents": 24
+    },
+    "pool/german/oasst2.jsonl": {
+      "sha256": "8ca93fad0812efdc21ed2ce06123627e97f41345eda8939419bd5af9b131ac01",
+      "bytes": 888635,
+      "documents": 410
+    },
+    "pool/german/ultrachat_de.jsonl": {
+      "sha256": "d42ccac07f7187c1f55f4de99335be4164f3b663d4319c2c29c31f234019d280",
+      "bytes": 1698445,
+      "documents": 488
+    },
+    "pool/german/wikipedia.jsonl": {
+      "sha256": "f0771f38f5f42aee0ff88741b7f6188b8b3bad3f0927181f1038716374285494",
+      "bytes": 2291623,
+      "documents": 443
+    },
+    "pool/graphics/legacy-dsv4.jsonl": {
+      "sha256": "09e62f79a8859e2c0f275bc38b93dde63f5751660eb3e9fc8636782cb47d3689",
+      "bytes": 2068241,
+      "documents": 335
+    },
+    "pool/graphics/repos.jsonl": {
+      "sha256": "862600d1842bd530e65884d8bc4bfcfa8892ddd1e430a927f30d35651539b7d1",
+      "bytes": 6231338,
+      "documents": 827
+    },
+    "pool/longctx/legacy-dsv4.jsonl": {
+      "sha256": "1f17bf4553c25558953fd746ce3bce278e932690e02d926395f8ecf21dacc100",
+      "bytes": 591648,
+      "documents": 10
+    },
+    "pool/longctx/module-bundles.jsonl": {
+      "sha256": "767b0cb154af75a8e5023e948f39fb3cbef424c99ba643779c5bd3e289945bb5",
+      "bytes": 9837165,
+      "documents": 83
+    },
+    "pool/multilingual/legacy-dsv4-wikipedia.jsonl": {
+      "sha256": "aa0344e84c60e1c423ac543d205bad77e3c0cc475799ddcb76d4e338cbdfb266",
+      "bytes": 420903,
+      "documents": 14
+    },
+    "pool/multilingual/wikipedia.jsonl": {
+      "sha256": "0cd11f4e6a768439f37aa354c422eb980d161e9ac2a8dde856a2f61a362480d0",
+      "bytes": 12207837,
+      "documents": 687
+    },
+    "pool/reasoning/synthetic/conversations.jsonl": {
+      "sha256": "9e890645baaad2045ac87ce9a9649082ccc37bec1c3d72b1da2a5b6c92af3e2f",
+      "bytes": 6724419,
+      "documents": 3015
+    },
+    "pool/reasoning/synthetic/legacy-dsv4-rendered.jsonl": {
+      "sha256": "e690d0acb22a427e0b17f57ee8f255a36c5550de23775a52c9117198b02e8f95",
+      "bytes": 440412,
+      "documents": 153
+    },
+    "pool/structured/legacy-dsv4.jsonl": {
+      "sha256": "e0a6537986934b5418b6d711c2712728979e1019eb6ca40885429fc068e8c6d4",
+      "bytes": 108168,
+      "documents": 19
+    },
+    "pool/structured/repos.jsonl": {
+      "sha256": "631c8fd66fa7fc1d60fe50e257385512c77e0c1d0e1a86f88a482c29e0b2f83c",
+      "bytes": 1226353,
+      "documents": 227
+    },
+    "pool/structured/synthetic/structured-in-chat.jsonl": {
+      "sha256": "435a1be5a7b8bd230105b85fbe708649fb2ddef5a65198bdd808bcff1b3d1b64",
+      "bytes": 395812,
+      "documents": 207
+    },
+    "pool/vocab_sweep/synthetic/kolibri-1-moe.jsonl": {
+      "sha256": "3d165633bb5201090c11a63671e4f993e054cc342881007eb03b29e03d89cf27",
+      "bytes": 1323571,
+      "documents": 110
+    },
+    "pool/vocab_sweep/synthetic/legacy-dsv4-rendered.jsonl": {
+      "sha256": "e30414d83fc63d1a5bb293a50d174eb65385efa237111f7ae001f923a9498424",
+      "bytes": 745316,
+      "documents": 65
+    },
+    "pool/vocab_sweep/synthetic/minicpm-v-4.6-moe.jsonl": {
+      "sha256": "c8bcca4781cb4b842f42e3c4c5c699effde6aca401e85f656ace5fd4946db2b4",
+      "bytes": 2509124,
+      "documents": 208
+    },
+    "pool/vocab_sweep/synthetic/minicpm-v-4.6.jsonl": {
+      "sha256": "c8bcca4781cb4b842f42e3c4c5c699effde6aca401e85f656ace5fd4946db2b4",
+      "bytes": 2509124,
+      "documents": 208
+    },
+    "pool/vocab_sweep/synthetic/minicpm-v-4.7-moe.jsonl": {
+      "sha256": "e2c3b5b513d5c43e4f57d04f2094b8d62ba775bf8559b2f116327bbc36321417",
+      "bytes": 2746220,
+      "documents": 216
+    },
+    "pool/vocab_sweep/synthetic/minicpm-v-4.7.jsonl": {
+      "sha256": "2dfd825f85e2c9f31a28e3d1dc1a1fe25b3fa06fd0b3ac94630d21130a3c0cec",
+      "bytes": 2512036,
+      "documents": 208
+    },
+    "pool/vocab_sweep/synthetic/muse-glimmer-30b.jsonl": {
+      "sha256": "4eb612024896df3b979fa74b92cc27ddbce56af100a14611bcfccd4ea9ce92c5",
+      "bytes": 2068343,
+      "documents": 172
+    },
+    "pool/vocab_sweep/synthetic/nemotron-3.5-lightning.jsonl": {
+      "sha256": "cd73220ef1ac82132a4c2b217f77af7b2eaea2c06939394fe36650221cc22a1d",
+      "bytes": 1269096,
+      "documents": 113
+    },
+    "pool/vocab_sweep/synthetic/nemotron-omni-moe.jsonl": {
+      "sha256": "2efb1a379e2e5b837a3baf362553e981b63fff67ca48762d093ebb91397b9e90",
+      "bytes": 1267229,
+      "documents": 113
+    },
+    "pool/vocab_sweep/synthetic/nemotron-puzzle-moe.jsonl": {
+      "sha256": "fd584ab5a8d7f8cd92208e1319697922955dcf914b452a8a356b41f7933a765e",
+      "bytes": 1267288,
+      "documents": 113
+    },
+    "pool/vocab_sweep/synthetic/qwen3.8-flash-next-moe.jsonl": {
+      "sha256": "ee4657dd2394309d06d6e6a978e020d535e0f899c0b80ca2067672889a276913",
+      "bytes": 2508961,
+      "documents": 208
+    }
+  },
+  "calib_longctx": {
+    "path": "calib_longctx.txt",
+    "sha256": "f04ebaf4030b3b5e4d785b4a7d326a907a053de7c4567f8d809bceca9a959007",
+    "documents": 29,
+    "tokens": 771448,
+    "document_tokens": {
+      "min": 17493,
+      "p50": 28668,
+      "max": 32706
+    },
+    "contiguous": true
+  }
+}

pool/vocab_sweep/synthetic/nemotron-omni-moe.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2efb1a379e2e5b837a3baf362553e981b63fff67ca48762d093ebb91397b9e90
+size 1267229

recipes/nemotron-omni-moe.yaml

@@ -0,0 +1,83 @@
+# Calibration recipe template for nemotron_h_moe models (Mamba-2 + attention + latent MoE): rendered
+# once for the rehearsal proxy NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B and again for Nemotron 3.5 Bolt
+# 74B-A9B if its template differs. node_corpus_nemotron.sh fills in name and repo.
+#
+# Against nemotron-3.5-lightning (calib-corpora, 2026-08-12):
+#   - same tokenizer (tokenizer.json, tokenizer_config.json, added tokens: fingerprints/), so the
+#     vocabulary sweep carries over;
+#   - Puzzle's template closes <think> with newlines on both sides and adds low_effort, so the
+#     Lightning build's markup is not this model's: re-rendered here (chat.format auto drives the
+#     model's own chat_template.jinja);
+#   - MoE shares instead of Lightning's dense-like ones.
+#
+# Architecture facts that matter here (Puzzle 75B-A9B, config.json):
+#   88 blocks: 40 Mamba-2, 40 MoE, 8 attention (5 of every 11 blocks carry state through Mamba-2 only)
+#   hidden 4096, latent 1024 (%256 = 0); expert width per layer 1280..2688 (2688 is not %256)
+#   512 routed experts, 4..18 active per layer: ~5M * 10 / 512 = ~100k tokens per expert
+#
+# Shares (they sum to 100):
+#   agentic 22      tool markup (<tool_call>, <tool_response>) is a dense token set natural text never emits
+#   code 16         the agentic benchmarks NVIDIA reports are code shaped
+#   reasoning 12    the template thinks by default; on MoE the slack goes to multilingual and vocab
+#   multilingual 18 non-Latin text routes to experts English never touches
+#   longctx 10      only 8 of 88 blocks are attention: long documents drive the Mamba state; plus calib_longctx
+#   vocab_sweep 12  every id routes somewhere
+#   structured 6    JSON, YAML, TOML, SQL, as files and inside assistant turns
+#   graphics 4
+
+name: nemotron-omni-moe
+
+model:
+  repo: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
+  tokenizer: /src/tokenizer.json
+  vocab_size: 131072
+  layers: 88
+  sliding_window: null
+  context_length: 262144
+  llama_cpp_commit: 'e64207bab6c2'
+
+seed: 20261015
+
+chat:
+  format: auto
+  # tokenizer_config.json has add_bos_token = false and the template emits no BOS (as Lightning)
+  add_bos_per_document: false
+
+calib_train:
+  target_tokens: 5000000
+  max_doc_fraction:
+    default: 0.05
+    longctx: 0.12
+  shares:
+    agentic: 22
+    code: 16
+    reasoning: 12
+    multilingual: 18
+    longctx: 10
+    vocab_sweep: 12
+    structured: 6
+    graphics: 4
+
+calib_longctx:
+  target_tokens: 750000
+  doc_tokens_min: 16384
+  doc_tokens_max: 32768
+  sources: [longctx]
+
+exclude:
+  # Anything pre-rendered with another model's markup, the Lightning (nemotron) render included:
+  # its </think> differs from this template's.
+  render: [dsv4, nemotron, muse-glimmer]
+  provenance_key: excluded_from_builds
+
+notes:
+  imatrix: >
+    llama-imatrix defaults to parse_special = false. Without --parse-special the chat markup in
+    calib_train.txt is tokenised as literal punctuation and the agentic and reasoning slices calibrate
+    on text the model never sees.
+  mtp: >
+    The MTP block is never executed by a plain forward pass: llama-imatrix collects it only with
+    --nextn. The profile (moe-nemotron-h) keeps it at q8_0, which needs no imatrix.
+  vision: >
+    Image and video markers are not in the corpus: text calibration cannot give the mmproj's
+    embedding slots honest statistics. corpus_check specials runs with them excepted.