replace with the locally built v1: 15 added/updated, 24 superseded removed

.gitattributes

@@ -33,3 +33,9 @@ code-daemon-relation-v1-s_win_x64_trt11.0_sm_80.engine filter=lfs diff=lfs merge
 code-daemon-relation-v1-l_linux_x64_trt11.0_sm_90.engine filter=lfs diff=lfs merge=lfs -text
 code-daemon-relation-v1-m_linux_x64_trt11.0_sm_90.engine filter=lfs diff=lfs merge=lfs -text
 code-daemon-relation-v1-s_linux_x64_trt11.0_sm_90.engine filter=lfs diff=lfs merge=lfs -text
+code-daemon-relation-v1-l_ov2026.4_cpu_fp16_b16_s256.bin filter=lfs diff=lfs merge=lfs -text
+code-daemon-relation-v1-l_ov2026.4_igpu_fp16_b16_s256.bin filter=lfs diff=lfs merge=lfs -text
+code-daemon-relation-v1-m_ov2026.4_cpu_fp16_b16_s128.bin filter=lfs diff=lfs merge=lfs -text
+code-daemon-relation-v1-m_ov2026.4_igpu_fp16_b16_s128.bin filter=lfs diff=lfs merge=lfs -text
+code-daemon-relation-v1-s_ov2026.4_cpu_fp16_b16_s64.bin filter=lfs diff=lfs merge=lfs -text
+code-daemon-relation-v1-s_ov2026.4_igpu_fp16_b16_s64.bin filter=lfs diff=lfs merge=lfs -text

README.md

@@ -42,9 +42,9 @@ class 0 is an explicit **abstain**, and a softmax threshold drops the rest of th
 
 | idx | label | meaning |
 |--:|---|---|
-| 0 | `NO_RELATION` | the two co-occur but are not related — **abstain** |
-| 1 | `semantically_similar_to` | near-duplicate purpose, or the same goal by a different mechanism |
-| 2 | `invalidates_with` | one supersedes, replaces or contradicts the other |
+| 0 | `NO_RELATION` | co-occur, unrelated — **abstain** |
+| 1 | `semantically_similar_to` | same purpose or same goal, other mechanism |
+| 2 | `supersedes_or_conflicts` | one supersedes, replaces or contradicts |
 | 3 | `depends_on` | one requires or configures the other |
 
 The taxonomy is deliberately **coarse**. An earlier 8-way version split these into near-synonym pairs
@@ -62,16 +62,13 @@ per-class maximum sags while "something is here" stays high.
 
 ## 2. Architecture
 
-| | |
-|---|---|
-| Warm-start | [`cross-encoder/mmarco-mMiniLMv2-L12-H384-v1`](https://huggingface.co/cross-encoder/mmarco-mMiniLMv2-L12-H384-v1) |
-| Encoder | XLM-RoBERTa, **12 layers / 384 hidden / 12 heads**, FFN 1536 |
-| Vocabulary | **250 006** SentencePiece pieces = 250 002 XLM-R + 4 marker tokens |
-| Markers | `[E1]` `[/E1]` `[E2]` `[/E2]` (ids 250002–250005) |
-| Sequence | **256 tokens** on the shipped engines (64 / 128 / 320 also provided) |
-| Inputs | `input_ids`, `attention_mask` — **no `token_type_ids`** |
-| Output | `logits[batch, 4]` |
-| Parameters | **~117M**, of which 96M is the multilingual embedding table |
+- **Warm-start** — [`cross-encoder/mmarco-mMiniLMv2-L12-H384-v1`](https://huggingface.co/cross-encoder/mmarco-mMiniLMv2-L12-H384-v1).
+- **Encoder** — XLM-RoBERTa, **12 layers / 384 hidden / 12 heads**, FFN 1536. **~117M**
+  parameters, of which 96M is the multilingual embedding table.
+- **Vocabulary** — **250 006** pieces = 250 002 XLM-R + 4 markers `[E1]` `[/E1]` `[E2]` `[/E2]`
+  (ids 250002–250005).
+- **I/O** — `input_ids`, `attention_mask` (**no `token_type_ids`**) → `logits[batch, 4]`.
+  Sequence **256** on the shipped engines; 64 / 128 also provided.
 
 ### Entity-marker pooling, not `[CLS]`
 
@@ -105,12 +102,16 @@ sequence budget is not a hard case, it is an unanswerable one.
 
 Measured on one laptop: Intel Core Ultra 9 275HX / NVIDIA RTX 5060 Laptop.
 
-| lane | batch × seq | per batch | throughput | per pair |
-|---|---|--:|--:|--:|
-| **TensorRT FP16, RTX 5060 Laptop** | 16 × 256 | **5.44 ms** | **2 942 pairs/s** | 0.34 ms |
-| **OpenVINO FP16, iGPU** (OV 2026.3) | 16 × 256 | **143 ms** | **111 pairs/s** | 8.96 ms |
-| OpenVINO FP16, CPU (OV 2026.3) | 16 × 256 | 366 ms | 44 pairs/s | 22.9 ms |
-| ONNX Runtime FP32, CPU | 16 × 256 | 380 ms | 42 pairs/s | 23.7 ms |
+All lanes at batch 16 × seq 256:
+
+| lane | pairs/s | per batch |
+|---|--:|--:|
+| **TensorRT FP16, RTX 5060** | **2 942** | 5.44 ms |
+| **OpenVINO FP16, iGPU** | **111** | 143 ms |
+| OpenVINO FP16, CPU | 44 | 366 ms |
+| ONNX Runtime FP32, CPU | 42 | 380 ms |
+
+OpenVINO rows were measured on 2026.3; the IRs in this repository are built for 2026.4 (re-timed on the embedding model: within a few percent either way).
 
 Per bucket, OpenVINO 2026.3 on the same laptop (pairs/s, solo):
 
@@ -119,7 +120,6 @@ Per bucket, OpenVINO 2026.3 on the same laptop (pairs/s, solo):
 | s | 16 × 64 | 159 | **508** |
 | m | 16 × 128 | 85 | **245** |
 | l | 16 × 256 | 44 | **111** |
-| xl | 16 × 320 | 34 | **85** |
 
 The integrated GPU is ~2.6× the CPU on every bucket, so on a host without a discrete card the
 iGPU lane is the one to route relation extraction to. Both devices at once give ~87 % of the sum
@@ -129,7 +129,7 @@ The GPU lane is ~67× the CPU lane, which is the point: relation extraction over
 of thousands of candidate pairs, and only the compiled-engine path makes that a background task
 rather than a batch job.
 
-Four **length buckets** ship — seq 64 / 128 / 256 / 320 at batch 16. Attention is quadratic in
+Three **length buckets** ship — seq 64 / 128 / 256 at batch 16. Attention is quadratic in
 sequence length, so routing short passages to a short engine is worth taking when your pairs vary in
 length. Padding is attention-masked, so a pair produces the same logits from any bucket that fits it.
 
@@ -138,10 +138,10 @@ length. Padding is attention-masked, so a pair produces the same logits from any
 The table above is a per-batch micro-benchmark. Sweeping a real corpus batches far wider, which
 changes what the bottleneck is:
 
-| | measured |
+| at corpus scale | measured |
 |---|--:|
-| GPU inference, 267 pairs per call at seq 256 | **5.87 ms** → ~45 000 pairs/s |
-| End-to-end incl. tokenisation (12 CPU threads, overlapped) | **~2 500 pairs/s** |
+| GPU inference, 267 pairs/call at seq 256 | **5.87 ms** → ~45 000 pairs/s |
+| End-to-end incl. tokenisation, 12 threads | **~2 500 pairs/s** |
 | Relations written, 3 043 chunks | 15 175 in 15.3 s → **994/s** |
 | Relations written, 4 975 chunks | 31 696 in 45.0 s → **704/s** |
 
@@ -151,13 +151,31 @@ what the length buckets are there to keep down.
 
 ---
 
+### Inside a real index
+
+Measured live in the UltraCode daemon over two repositories, TensorRT FP16, wide batches of
+~263 pairs per call:
+
+| metric | value |
+|---|--:|
+| pairs scored | **2 162 / s** sustained (210 071 in 97.2 s) |
+| inference share of wall | 90 % |
+| padding wasted | 33 % |
+| relations created | 1 011 – 1 345 / s |
+
+The live rate is below the 2 942 pairs/s micro-benchmark above because a third of every wide
+batch is padding. Relations *created* run lower still: most scored pairs fall under the
+acceptance threshold and are rejected, which is the point of the threshold.
+
+---
+
 ## 5. Standalone use
 
 ```python
 import numpy as np, onnxruntime as ort
 from transformers import AutoTokenizer
 
-LABELS = ["NO_RELATION", "semantically_similar_to", "invalidates_with", "depends_on"]
+LABELS = ["NO_RELATION", "semantically_similar_to", "supersedes_or_conflicts", "depends_on"]
 
 tok  = AutoTokenizer.from_pretrained(".")        # includes the [E1]/[E2] marker tokens
 sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
@@ -174,7 +192,7 @@ def classify(marked_text, max_len=256, tau=0.7):
     return (LABELS[i], float(p[i])) if i else ("NO_RELATION", float(p[0]))
 
 classify("The [E1]FAISS[/E1] index was replaced by the [E2]native IVF[/E2] backend.")
-# -> ('invalidates_with', 0.7x)
+# -> ('supersedes_or_conflicts', 0.7x)
 ```
 
 Both entities must appear **inside one passage**, marked in place. The model reads context, so a bare
@@ -193,6 +211,45 @@ class plus the `tau` gate exist because the useful operating point is *high-prec
 maximum recall: for building a graph, the real test is spot-checking the edges it emits at your
 chosen threshold.
 
+### Against zero-shot NLI — the alternative to training a head
+
+With no training data for these classes, the textbook approach is a **zero-shot NLI** model: state
+each class as a hypothesis and take the entailment score.
+[`MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7`](https://huggingface.co/MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7)
+is the standard pick. Same 691 rows, same 4-way mapping, same machine.
+
+**What matches and what does not.** There is no public model trained on these four merged
+classes, so the comparator matches on **task and dev set** but not on size (279 M against 117 M)
+or runtime (PyTorch against onnxruntime) — and zero-shot pays one forward pass per class by
+construction, which the speed figure below makes visible:
+
+| model | params | macro-F1 | accuracy |
+|---|--:|--:|--:|
+| **this model** | 117 M | **0.569** | **0.670** |
+| mDeBERTa-v3-xnli, zero-shot | 279 M | 0.194 | 0.295 |
+| always `NO_RELATION` | — | 0.158 | 0.463 |
+
+On the question the daemon actually asks — *is there a relation at all* — the gap is narrower on
+F1 and wider in kind:
+
+| model | has-rel F1 | P | R |
+|---|--:|--:|--:|
+| **this model** | **0.810** | **0.793** | 0.828 |
+| mDeBERTa-v3-xnli | 0.699 | 0.538 | **0.997** |
+
+A recall of 0.997 at a precision of 0.538 is the whole story: asked whether two entities in the
+same paragraph are related, an NLI model says yes almost every time. For a graph builder that reads
+the score as an edge weight, that is not a usable signal.
+
+**Speed, same hardware, same 691 pairs:** **351 pairs/s** for this model (onnxruntime CUDA, FP32,
+batch 16 × seq 256) against **61 pairs/s** for the zero-shot baseline (PyTorch CUDA). The 5.8× is
+structural, not an implementation detail — zero-shot needs one forward pass **per class**, so a
+4-class taxonomy costs four passes per pair where this model costs one. The shipped TensorRT FP16
+engine is faster again (2 942 pairs/s, above).
+
+Measured 2026-09-11; harness and raw JSON ship in the UltraCode repo
+(`models/_distill_shared/bench_vs_generic.py`).
+
 **Suited to**
 - Turning prose or documentation into a typed concept graph.
 - Any sweep where a large LLM per pair would be too slow or too expensive.
@@ -209,9 +266,10 @@ chosen threshold.
 
 Compiled engines, named per **runtime × OS × GPU arch**, plus the ONNX for standalone use.
 
-- **TensorRT FP16** — `code-daemon-relation-v1-{s,m,l,xl}_{win_x64,linux_x64}_trt11.0_sm_120.engine`
-  (buckets seq 64 / 128 / 256 / 320, batch 16).
-- **OpenVINO FP16** — `code-daemon-relation-v1-{s,m,l,xl}_ov2026.3_{cpu,igpu}_fp16_b16_s{64,128,256,320}.{xml,bin}`.
+- **TensorRT FP16** — `code-daemon-relation-v1-{s,m,l}_{win_x64,linux_x64}_trt11.0_sm_{75,80,86,89,120}.engine`,
+  plus `code-daemon-relation-v1-{s,m,l}_linux_x64_trt11.0_sm_90.engine` (H100 / H200, Linux only)
+  (buckets seq 64 / 128 / 256, batch 16).
+- **OpenVINO FP16** — `code-daemon-relation-v1-{s,m,l}_ov2026.4_{cpu,igpu}_fp16_b16_s{64,128,256}.{xml,bin}`.
 - **Tokenizer** — `tokenizer.json`, `sentencepiece.bpe.model`, `tokenizer_config.json` (XLM-R
   SentencePiece with the four marker tokens added).
 - **ONNX** — `model.onnx` (+ `model.onnx.data`), FP32, the build source for every engine above.
@@ -225,9 +283,8 @@ lossy, and FP16 costs nothing on any GPU that can run it.
 
 Released **MIT**.
 
-| Source | Note |
-|---|---|
-| `cross-encoder/mmarco-mMiniLMv2-L12-H384-v1` (warm-start) | **mMARCO ← MS MARCO → non-commercial research terms** |
+**Warm-start base:** `cross-encoder/mmarco-mMiniLMv2-L12-H384-v1` — mMARCO ← MS MARCO, whose terms
+are **non-commercial research**.
 
 ⚠️ The warm-start base derives from **MS MARCO (non-commercial)**. Whether a fine-tuned model
 inherits dataset-use terms is legally unsettled — this is **not legal advice**. Retrain from a

code-daemon-relation-v1-l_ov2026.3_cpu_fp16_b16_s256.xml → code-daemon-relation-v1-l_ov2026.4_cpu_fp16_b16_s256.xml

@@ -21602,6 +21602,6 @@
 		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
 	</edges>
 	<rt_info>
-		<info name="OpenVINO Runtime" value="2026.3.0-22451-8a17657b995-releases/2026/3" />
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
 	</rt_info>
 </net>

code-daemon-relation-v1-l_ov2026.3_igpu_fp16_b16_s256.xml → code-daemon-relation-v1-l_ov2026.4_igpu_fp16_b16_s256.xml

@@ -21602,6 +21602,6 @@
 		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
 	</edges>
 	<rt_info>
-		<info name="OpenVINO Runtime" value="2026.3.0-22451-8a17657b995-releases/2026/3" />
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
 	</rt_info>
 </net>

code-daemon-relation-v1-m_ov2026.3_cpu_fp16_b16_s128.xml → code-daemon-relation-v1-m_ov2026.4_cpu_fp16_b16_s128.xml

@@ -21602,6 +21602,6 @@
 		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
 	</edges>
 	<rt_info>
-		<info name="OpenVINO Runtime" value="2026.3.0-22451-8a17657b995-releases/2026/3" />
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
 	</rt_info>
 </net>

code-daemon-relation-v1-m_ov2026.3_igpu_fp16_b16_s128.xml → code-daemon-relation-v1-m_ov2026.4_igpu_fp16_b16_s128.xml

@@ -21602,6 +21602,6 @@
 		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
 	</edges>
 	<rt_info>
-		<info name="OpenVINO Runtime" value="2026.3.0-22451-8a17657b995-releases/2026/3" />
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
 	</rt_info>
 </net>

code-daemon-relation-v1-s_ov2026.3_cpu_fp16_b16_s64.xml → code-daemon-relation-v1-s_ov2026.4_cpu_fp16_b16_s64.xml

@@ -21602,6 +21602,6 @@
 		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
 	</edges>
 	<rt_info>
-		<info name="OpenVINO Runtime" value="2026.3.0-22451-8a17657b995-releases/2026/3" />
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
 	</rt_info>
 </net>

code-daemon-relation-v1-s_ov2026.3_igpu_fp16_b16_s64.xml → code-daemon-relation-v1-s_ov2026.4_igpu_fp16_b16_s64.xml

@@ -21602,6 +21602,6 @@
 		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
 	</edges>
 	<rt_info>
-		<info name="OpenVINO Runtime" value="2026.3.0-22451-8a17657b995-releases/2026/3" />
+		<info name="OpenVINO Runtime" value="2026.4.0-22959-99c81491cc3-releases/2026/4" />
 	</rt_info>
 </net>

code-daemon-relation-v1-xl_linux_x64_trt11.0_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3e200a40f8dbf2c7c812a50c5bc5710f734c5f869d3bdb8b3364879b9ef361e3
-size 238387732

code-daemon-relation-v1-xl_linux_x64_trt11.0_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1b48851f854a7f3f62657735e98e6063088eae357e245e67db44440b8a52efd7
-size 237937980

code-daemon-relation-v1-xl_linux_x64_trt11.0_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:eabd366dd653374a1325ef603a2737e08b222a2654495b8f84279bd6f4d2c840
-size 238309732

code-daemon-relation-v1-xl_linux_x64_trt11.0_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:983ba8d9a7c73906b28410223a7cf2b2987672ac4f848858a3957c958d241dae
-size 239674964

code-daemon-relation-v1-xl_ov2026.3_cpu_fp16_b16_s320.bin

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

code-daemon-relation-v1-xl_ov2026.3_igpu_fp16_b16_s320.bin

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

code-daemon-relation-v1-xl_win_x64_trt11.0_sm_120.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4d83141e50053725d59d49fa6074c1be0db9cce9eace64bd3dceeebfa41961f7
-size 240319756

code-daemon-relation-v1-xl_win_x64_trt11.0_sm_75.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3456ad7f4a1f1b7f74bf128789537b4ea649cfd21f76b247fc3c46abcb1aa35c
-size 237830228

code-daemon-relation-v1-xl_win_x64_trt11.0_sm_86.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3d459caa6d37cb81b1c5ebd33284e182099646cca8845286ae58fcab19c0ca76
-size 239786044

code-daemon-relation-v1-xl_win_x64_trt11.0_sm_89.engine

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bd69f48dcd93bc6a41a475091d41cb1ad1479f5268600a37bc915ebb75b8243a
-size 236747476

manifest.json

@@ -9,10 +9,10 @@
   ],
   "runtime_tags": {
     "trt": "trt11.0",
-    "ov": "ov2026.3",
+    "ov": "ov2026.4",
     "tvm": "tvm0.25",
     "mlx_dir": "model_gpu_mlx0.22"
   },
-  "compiled_at": "2026-08-05T08:55:19Z",
+  "compiled_at": "2026-09-16T20:05:04Z",
   "compiled_by": "models/_compile"
 }

model_gpu_mlx0.22/config.json

@@ -1,50 +1,45 @@
-{
-  "add_cross_attention": false,
-  "architectures": [
-    "XLMRobertaForSequenceClassification"
-  ],
-  "attention_probs_dropout_prob": 0.1,
-  "bos_token_id": 0,
-  "classifier_dropout": null,
-  "dtype": "float32",
-  "eos_token_id": 2,
-  "hidden_act": "gelu",
-  "hidden_dropout_prob": 0.1,
-  "hidden_size": 384,
-  "id2label": {
-    "0": "NO_RELATION",
-    "1": "semantically_similar_to",
-    "2": "shares_purpose_with",
-    "3": "invalidates_with",
-    "4": "configured_by",
-    "5": "depends_on",
-    "6": "contradicts",
-    "7": "replaced_by"
-  },
-  "initializer_range": 0.02,
-  "intermediate_size": 1536,
-  "is_decoder": false,
-  "label2id": {
-    "NO_RELATION": 0,
-    "configured_by": 4,
-    "contradicts": 6,
-    "depends_on": 5,
-    "invalidates_with": 3,
-    "replaced_by": 7,
-    "semantically_similar_to": 1,
-    "shares_purpose_with": 2
-  },
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 514,
-  "model_type": "xlm-roberta",
-  "num_attention_heads": 12,
-  "num_hidden_layers": 12,
-  "pad_token_id": 1,
-  "position_embedding_type": "absolute",
-  "sbert_ce_default_activation_function": "torch.nn.modules.linear.Identity",
-  "tie_word_embeddings": true,
-  "transformers_version": "5.1.0",
-  "type_vocab_size": 1,
-  "use_cache": true,
-  "vocab_size": 250006
-}
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "XLMRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "custom_head": "entity_start_marker_meanmask (concat E1,E2 -> Linear); [CLS] is not read",
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "id2label": {
+    "0": "NO_RELATION",
+    "1": "semantically_similar_to",
+    "2": "invalidates_with",
+    "3": "depends_on"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "is_decoder": false,
+  "label2id": {
+    "NO_RELATION": 0,
+    "depends_on": 3,
+    "invalidates_with": 2,
+    "semantically_similar_to": 1
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "xlm-roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "num_labels": 4,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "sbert_ce_default_activation_function": "torch.nn.modules.linear.Identity",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.1.0",
+  "type_vocab_size": 1,
+  "ultracode_note": "Inference artefact is model.onnx (inputs: input_ids, attention_mask; output: logits[batch, 4]). Feed the pair as (empty query, marked passage).",
+  "use_cache": true,
+  "vocab_size": 250006
+}