raw float weights: model.safetensors + config.json, exported from the shipped model.onnx

.gitattributes

@@ -41,3 +41,4 @@ code-daemon-relation-v1-s_ov2026.4_cpu_fp16_b16_s64.bin filter=lfs diff=lfs merg
 code-daemon-relation-v1-s_ov2026.4_igpu_fp16_b16_s64.bin filter=lfs diff=lfs merge=lfs -text
 model.onnx filter=lfs diff=lfs merge=lfs -text
 model.onnx.data filter=lfs diff=lfs merge=lfs -text
+model.safetensors filter=lfs diff=lfs merge=lfs -text

README.md

@@ -273,6 +273,14 @@ Compiled engines, named per **runtime × OS × GPU arch**, plus the ONNX for sta
 - **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.
+- **Raw weights** — `model.safetensors` + `config.json`, the same FP32 weights under transformers
+  names: the encoder as `roberta.*`, the head as `classifier.weight` `[4, 768]` + `classifier.bias`.
+  ⚠ The head is NOT the stock one — `AutoModelForSequenceClassification` would leave its own
+  `classifier.dense` / `classifier.out_proj` randomly initialised and ignore ours. 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`, 250002 and 250004), concatenate the two and
+  apply the classifier. Done that way it matches the ONNX to 2e-6. There is no MLX build of this
+  model: the Apple runtime implements no entity-pool head.
 
 FP16 rather than INT8: this architecture's activation outliers make per-tensor INT8 calibration
 lossy, and FP16 costs nothing on any GPU that can run it.

config.json

@@ -15,7 +15,7 @@
   "id2label": {
     "0": "NO_RELATION",
     "1": "semantically_similar_to",
-    "2": "invalidates_with",
+    "2": "supersedes_or_conflicts",
     "3": "depends_on"
   },
   "initializer_range": 0.02,
@@ -23,9 +23,9 @@
   "is_decoder": false,
   "label2id": {
     "NO_RELATION": 0,
-    "depends_on": 3,
-    "invalidates_with": 2,
-    "semantically_similar_to": 1
+    "semantically_similar_to": 1,
+    "supersedes_or_conflicts": 2,
+    "depends_on": 3
   },
   "layer_norm_eps": 1e-05,
   "max_position_embeddings": 514,
@@ -41,5 +41,17 @@
   "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
+  "vocab_size": 250006,
+  "ultracode": {
+    "format": "raw-safetensors/1",
+    "exported_from": "model.onnx",
+    "onnx_sha256": "73556d636633563bd0688d52899c475ff61a48f0687e8c48703ee4b9bc74d135",
+    "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",
+    "entity_marker_ids": [
+      250002,
+      250004
+    ]
+  }
 }

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43e335634a22eb24bbc71f10c5ec26af62a53b8605187edc9208d739659c2a8a
+size 470013960