compiled artefacts from win_x64 (sm_120)

README.md

@@ -14,7 +14,7 @@ tags:
 - openvino
 pipeline_tag: text-classification
 base_model:
-- cross-encoder/mmarco-mMiniLMv2-L12-H384-v1
+- hotchpotch/mMiniLMv2-L12-H384
 ---
 
 # code-daemon-relation-v1
@@ -29,6 +29,11 @@ sweep a whole corpus — **~2 900 pairs/s** on a laptop RTX 5060 — and its hea
 marked entities directly instead of a whole-passage summary, so the relation is scored between
 them and in order. Multilingual: the XLM-R backbone reads prose and code comments in many languages.
 
+> The weights under this id were replaced on 2026-10-12. Same architecture, same inputs and
+> outputs, same tokenizer; the backbone is now the MIT-licensed
+> `hotchpotch/mMiniLMv2-L12-H384`, and macro-F1 on the held-out set moved from 0.57 to 0.60.
+> Engines built for the previous weights are gone from this repository.
+
 ## The four classes
 
 Wrap the entities in `[E1]…[/E1]` and `[E2]…[/E2]` in their own context.
@@ -54,7 +59,7 @@ training data is a zero-shot NLI model that scores each class as a hypothesis:
 
 | model | params | macro-F1 | "has a relation" F1 | P | R | pairs/s |
 |---|--:|--:|--:|--:|--:|--:|
-| **this model** | 117M | **0.569** | **0.810** | **0.793** | 0.828 | **351** |
+| **this model** | 117M | **0.598** | **0.798** | **0.786** | 0.811 | **351** |
 | mDeBERTa-v3-xnli, zero-shot | 279M | 0.194 | 0.699 | 0.538 | **0.997** | 61 |
 | always `NO_RELATION` | — | 0.158 | — | — | — | — |
 
@@ -63,9 +68,11 @@ training data is a zero-shot NLI model that scores each class as a hypothesis:
 - **5.8× faster on the same runtime, structurally**: zero-shot pays one forward pass per class,
   this pays one per pair. Speeds above are onnxruntime CUDA FP32 at batch 16 × seq 256; the
   shipped TensorRT engine is faster again.
-- **The trade:** three relation types plus abstain, nothing finer, and macro-F1 0.57 — the
-  classes are imbalanced and some boundaries are genuinely ambiguous. It is built for a
-  high-precision operating point: spot-check the edges at the `tau` you choose.
+- **The trade:** three relation types plus abstain, nothing finer, and macro-F1 0.60 — the
+  classes are imbalanced and some boundaries are genuinely ambiguous (per class on the same set:
+  `NO_RELATION` 0.76, `semantically_similar_to` 0.64, `supersedes_or_conflicts` 0.47,
+  `depends_on` 0.52). It is built for a high-precision operating point: spot-check the edges at
+  the `tau` you choose.
 
 ### Speed by lane
 
@@ -122,13 +129,15 @@ already have); entities more than a 256-token window apart.
 
 ## How it was made
 
-- **Warm start:** the multilingual ranking cross-encoder `mmarco-mMiniLMv2-L12-H384-v1`
-  (XLM-RoBERTa, 12 layers × 384), with four marker tokens added to its vocabulary.
-- **New head:** its single ranking logit was replaced by a 4-class layer over the hidden states at
-  the `[E1]` and `[E2]` positions, concatenated — not over `[CLS]`. Direction comes free: swap
-  the markers and the input genuinely changes.
+- **Backbone:** `hotchpotch/mMiniLMv2-L12-H384` — Microsoft's multilingual MiniLM v2 (12 layers ×
+  384, distilled from XLM-RoBERTa Large), a plain masked-language-model encoder with no
+  task-specific stage in front of it — with four marker tokens added to its vocabulary.
+- **Head:** a 4-class layer over the hidden states at the `[E1]` and `[E2]` positions,
+  concatenated — not over `[CLS]`. Direction comes free: swap the markers and the input genuinely
+  changes.
 - **Training:** sequence-level distillation from an LLM teacher on relation tuples, each grounded
-  in its passage; `NO_RELATION` is a trained class, not an inferred one.
+  in its passage, with a class-balanced (focal) objective; `NO_RELATION` is a trained class, not
+  an inferred one.
 - **FP16 engines**, not INT8: this architecture's activation outliers make INT8 calibration lossy.
 
 ## Files
@@ -138,24 +147,21 @@ already have); entities more than a 256-token window apart.
 | `model.onnx` + `model.onnx.data` | FP32 ONNX — the build source and the standalone path |
 | `model.safetensors`, `config.json` | the same weights: encoder as `roberta.*`, head as `classifier.weight` `[4, 768]` + `classifier.bias` (see below); the Apple (MLX) build is prepared from this pair |
 | `tokenizer.json`, `tokenizer_config.json`, `sentencepiece.bpe.model` | XLM-R tokenizer with the four marker tokens |
-| `code-daemon-relation-v1-{s,m,l}_{win_x64,linux_x64}_trt11.0_sm_{75,80,86,89,120}.engine` | TensorRT FP16 per bucket × OS × GPU; `sm_90` (H100/H200) Linux only |
+| `code-daemon-relation-v1-{s,m,l}_{win_x64,linux_x64}_trt11.3_sm_{75,80,86,89,90,120}.engine` | TensorRT FP16 per bucket × OS × GPU; `sm_90` (H100/H200) Linux only. The set is filled per architecture as each is built — the file list is the truth |
 | `code-daemon-relation-v1-{s,m,l}_ov2026.4_{cpu,igpu}_fp16_b16_s{64,128,256}.{xml,bin}` | OpenVINO 2026.4 FP16, Intel CPU / iGPU |
 | `coreml_ane/embed.mlpackage/` | Core ML for the Apple Neural Engine, functions `b16 s64 / s128 / s256`; load with `cpuAndNeuralEngine` |
 
 ⚠ The head is not the stock `transformers` one — `AutoModelForSequenceClassification` would
 initialise its own and ignore this. Load the encoder with `AutoModel`, mean-pool the last hidden
 state over the tokens equal to each marker id (`config.json` → `ultracode.entity_marker_ids`),
-concatenate the two and apply the classifier; that matches the ONNX to 2e-6.
+concatenate the two and apply the classifier; that matches the ONNX to 3e-6.
 
 ## License & attribution
 
 Released under the **MIT license**.
 
-⚠️ The warm-start base `cross-encoder/mmarco-mMiniLMv2-L12-H384-v1` derives from mMARCO ← MS MARCO,
-whose terms are **non-commercial research**. Whether a fine-tuned model inherits dataset-use terms
-is legally unsettled — this is **not legal advice**. Retrain from a permissive base if strict
-compliance matters to you.
-
-Warm-started from **[cross-encoder/mmarco-mMiniLMv2-L12-H384-v1](https://huggingface.co/cross-encoder/mmarco-mMiniLMv2-L12-H384-v1)**.
-Backbone: XLM-RoBERTa. Used by the [Code-Daemon](https://code-daemon.dev) code assistant,
-though nothing about the model is specific to it.
+Backbone: **[hotchpotch/mMiniLMv2-L12-H384](https://huggingface.co/hotchpotch/mMiniLMv2-L12-H384)**
+(MIT), Microsoft's
+[Multilingual-MiniLM-L12-H384](https://github.com/microsoft/unilm/tree/master/minilm) distilled
+from XLM-RoBERTa Large. Used by the [Code-Daemon](https://code-daemon.dev) code assistant, though
+nothing about the model is specific to it.

code-daemon-relation-v1-l_linux_x64_trt11.3_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:faf76282fcfd85091e95790f29e06968980ba7fe16c37e8a8a2356102fab4cff
-size 238305756

code-daemon-relation-v1-l_linux_x64_trt11.3_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:74fb4e2dad41ee877940b4abac3d7be4fd21e519e44d889294044b88ad27c17a
-size 237255908

code-daemon-relation-v1-l_linux_x64_trt11.3_sm_80.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8e255e35df614657f14af1ad9e431f23c6c80663c35cea0b2f7c2b64035c1680
-size 239758380

code-daemon-relation-v1-l_linux_x64_trt11.3_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c5f4b2816e8b922be019307d87daa2d061750a5c7bac56d4d700c761faf6e38d
-size 240116716

code-daemon-relation-v1-l_linux_x64_trt11.3_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7e30983a2a6e996c05492804f36f35189a2b129ef905e9b23c624a673dd627e8
-size 240183756

code-daemon-relation-v1-l_linux_x64_trt11.3_sm_90.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bcaa26b65ea254cc7e756aa85601192ea7bf56247a62300a407fc1d72ab31a72
-size 237340532

code-daemon-relation-v1-l_ov2026.4_cpu_fp16_b16_s256.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0c5212f004254b5de637be1b845bd77cf27ce28b9faa6116080ee1000788a63b
-size 234999234

code-daemon-relation-v1-l_ov2026.4_igpu_fp16_b16_s256.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0c5212f004254b5de637be1b845bd77cf27ce28b9faa6116080ee1000788a63b
-size 234999234

code-daemon-relation-v1-l_win_x64_trt11.3_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bbd8f86bed1bb3e1445d7decda7bc74a2b3f9d1466ce10bea8bb03094dec0b65
-size 240345556

code-daemon-relation-v1-l_win_x64_trt11.3_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:cbea2ea7d4a6a5ce03413f311e66ebe6e21e57c93d495e172aa3ee103a63ef5b
-size 238443276

code-daemon-relation-v1-l_win_x64_trt11.3_sm_80.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b083f76c5511b7ce85ea5453190c2db2630d29966c15a39fe49cfba25486c9d8
-size 238403444

code-daemon-relation-v1-l_win_x64_trt11.3_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:03e8029c976f8e8dadbd426d8b86bf6f9d94c58e1a54702b752fd072704a5e56
-size 238908652

code-daemon-relation-v1-l_win_x64_trt11.3_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:44412683abbdeb03e15b2fc1895cd1a845f4e1f77587f45c3e95877fe75b3025
-size 237221100

code-daemon-relation-v1-m_linux_x64_trt11.3_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:fc59ce0eb014d452478bc96da9a3a4ce5388c098362e884925046ffeb3098b1c
-size 238398708

code-daemon-relation-v1-m_linux_x64_trt11.3_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:56e09c7f363164280d96169427451887ba81473f96b874b2edd3c0caa87a93a9
-size 237200580

code-daemon-relation-v1-m_linux_x64_trt11.3_sm_80.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6e1a39e0b4cbfaafeb3e96c06e3b4df2688ba0df133150915cbb193bd3758bb4
-size 240150460

code-daemon-relation-v1-m_linux_x64_trt11.3_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5d4871c8752464bb09441df07a96545ad24d226a0ed2bce487ce2624580ce552
-size 240059644

code-daemon-relation-v1-m_linux_x64_trt11.3_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f9d8288d7f19890a477338ec09e66997c70f0420fa92b7e264ed1177298b8a60
-size 238786980

code-daemon-relation-v1-m_linux_x64_trt11.3_sm_90.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8265bb8d92b8def962e6ff034f745e43997365e41abc3cb9b852dde3ceb930e1
-size 237319724

code-daemon-relation-v1-m_ov2026.4_cpu_fp16_b16_s128.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0c5212f004254b5de637be1b845bd77cf27ce28b9faa6116080ee1000788a63b
-size 234999234

code-daemon-relation-v1-m_ov2026.4_igpu_fp16_b16_s128.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0c5212f004254b5de637be1b845bd77cf27ce28b9faa6116080ee1000788a63b
-size 234999234

code-daemon-relation-v1-m_win_x64_trt11.3_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f9ae683602f90a6ec769fb1e224f0ba39e2f158aa97469aef6303a78f6a9b9ce
-size 240317492

code-daemon-relation-v1-m_win_x64_trt11.3_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:49589562a3c528bf01012b5059d1ca2c8967e9e55402a7cd9eeadfcf49e8e923
-size 237181924

code-daemon-relation-v1-m_win_x64_trt11.3_sm_80.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6692fb1fde6a78f7779f9942d6fa910ee49d388e73e6fc42cd6bc19d8a2d420d
-size 238279020

code-daemon-relation-v1-m_win_x64_trt11.3_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:da6b3d802dc9455633c4d6378adc59c3fcc5b6c6c4c02c453ff1d4275d3b67e7
-size 238424404

code-daemon-relation-v1-m_win_x64_trt11.3_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7da90f1252b99a7d949abef50b1ee62e148a4331cbac9b0f677e1937f07ecf6b
-size 237224140

code-daemon-relation-v1-s_linux_x64_trt11.3_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:63130f4245adb15a5c917c1e4008f3f4cee009e28114d65abf17e795f3daf2db
-size 238332868

code-daemon-relation-v1-s_linux_x64_trt11.3_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:54108ccfb78f6a2af4e54c79978130d89f36c6786d2be9b8b885d832fe4877b9
-size 238288932

code-daemon-relation-v1-s_linux_x64_trt11.3_sm_80.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d10767f2ad75192500ce9216bbe35b9f6e8e705138374d46266361db8e395d95
-size 239500116

code-daemon-relation-v1-s_linux_x64_trt11.3_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a12a4464d7e45d5e31c16f0ea26e5b581ebd25d5768de318451f6816d26597e0
-size 239981500

code-daemon-relation-v1-s_linux_x64_trt11.3_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3463780eb86e564b809cdabd9a0cf61a5c5ddeaa17670d70af1488062bc6f3fc
-size 238663060

code-daemon-relation-v1-s_linux_x64_trt11.3_sm_90.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:231d9f3bba7955e8b625bc66d16a93ac54a7c2115688654388caa844c33b2c4f
-size 237151924

code-daemon-relation-v1-s_ov2026.4_cpu_fp16_b16_s64.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0c5212f004254b5de637be1b845bd77cf27ce28b9faa6116080ee1000788a63b
-size 234999234

code-daemon-relation-v1-s_ov2026.4_igpu_fp16_b16_s64.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0c5212f004254b5de637be1b845bd77cf27ce28b9faa6116080ee1000788a63b
-size 234999234

code-daemon-relation-v1-s_win_x64_trt11.3_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6a18709f8ae6e6c97c5f1fb4b88a3eac1699a2c9443b81e24ef3de5ed06f8c51
-size 238321548

code-daemon-relation-v1-s_win_x64_trt11.3_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:43c0ef8d8c46ee944422615f2a1e0ef20db4324077db6c3723427edee34bee67
-size 237063204

code-daemon-relation-v1-s_win_x64_trt11.3_sm_80.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:24c4762fa4dcc34da56822641abc248c739b0fb5f96b279a9293fc3c83c6f270
-size 238613724

code-daemon-relation-v1-s_win_x64_trt11.3_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3905c36adfad6499d7c5337de36f1f2f3b9a76d6b3cb62d3003f101bee308bf8
-size 238579844

code-daemon-relation-v1-s_win_x64_trt11.3_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5bcb70ebda846a552e4f835666bd727f7e0615764cb234478beaeec3837aaebe
-size 237082452

config.json

@@ -45,7 +45,7 @@
   "ultracode": {
     "format": "raw-safetensors/1",
     "exported_from": "model.onnx",
-    "onnx_sha256": "73556d636633563bd0688d52899c475ff61a48f0687e8c48703ee4b9bc74d135",
+    "onnx_sha256": "9189bfedd376d2a08acd18e7b490000d5f78458c3e4388d1089704319a115840",
     "head": "entity_pool",
     "position_ids": "roberta",
     "pooling": "mean of the last hidden state over the tokens equal to each marker id; concat(E1, E2) -> classifier",

manifest.json

@@ -1,17 +0,0 @@
-{
-  "model_id": "code-daemon-relation-v1",
-  "dimension": 4,
-  "max_tokens": 256,
-  "quantization": "fp16",
-  "targets": [
-    "openvino"
-  ],
-  "runtime_tags": {
-    "trt": "trt11.3",
-    "ov": "ov2026.4",
-    "tvm": "tvm0.25",
-    "mlx_dir": "model_gpu_mlx0.32"
-  },
-  "compiled_at": "2026-09-24T02:28:50Z",
-  "compiled_by": "models/_compile"
-}

model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:73556d636633563bd0688d52899c475ff61a48f0687e8c48703ee4b9bc74d135
-size 1548544
+oid sha256:9189bfedd376d2a08acd18e7b490000d5f78458c3e4388d1089704319a115840
+size 1558814

model.onnx.data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b51662317655847207a9b2c146d7a5084310fac4860e7169cd47c1c161e25308
+oid sha256:966720aa7a9edbcef00bbd961e2fb61d3a616e460f21cc4cd2de73b058a9cc73
 size 470057984