GLiNER2.5-Decide Core ML: fp16/W8/LUT6 at L128, L256, L512 (mobius 01e6e6a)

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -0,0 +1,95 @@
+---
+license: apache-2.0
+library_name: coremltools
+pipeline_tag: text-classification
+base_model: fastino/GLiNER2.5-Decide
+tags:
+- coreml
+- gliner2
+- apple-silicon
+- classification
+---
+
+# GLiNER2.5-Decide for Core ML
+
+Fixed-shape Core ML exports of the classification path of [fastino/GLiNER2.5-Decide](https://huggingface.co/fastino/GLiNER2.5-Decide) at revision `65624f1a0265b3f612bae66a2685a06b94a68a9d`. Fastino authored the Apache-2.0 source model; Fluid Inference converted it. Each package holds the DeBERTa-v3-large encoder and the shared label classifier (436,022,273 exported parameters). The span and count heads are not exported: this release covers classification only.
+
+One call scores up to four decision heads (for example intent, urgency and route) over the same text. The host tokenizes the schema with the upstream GLiNER2 processor and passes the position of every label marker; the package returns per-head logits. Single-label heads use softmax and multi-label heads use sigmoid with `cls_threshold`, exactly as `classify_text` does.
+
+## Fast Decisions
+
+Scored on Fastino's [Fast Decisions](https://huggingface.co/datasets/fastino/fast-decisions) development split (revision `1a33070c`, 17 domains × 100 rows, 2,900 heads) with the protocol from the dataset card: one call per head with the labels as given, exact set match, mean of the 17 domain accuracies. The 512-token bucket fits every row.
+
+| Package (L512) | Size | Average | Pooled | Heads changed vs native |
+| --- | ---: | ---: | ---: | ---: |
+| native PyTorch | — | 62.93% | 61.38% | — |
+| `fp16` | 936 MB | 62.93% | 61.38% | 0 |
+| `w8` | 477 MB | 63.01% | 61.52% | 7 |
+| `lut6` | 361 MB | 63.03% | 61.52% | 42 |
+
+The card's published 60.2% for GLiNER2.5-Decide is on the held-out test split (300 rows per domain), which is not public; the card asks that development-split scores not be reported as the benchmark. The table above shows that the Core ML packages reproduce the native checkpoint on the public split, not a new benchmark result.
+
+Rejected variants, kept in `reports/` for reference: embedding-only W8 (805 MB, 63.00%) saves little over `w8`; per-tensor LUT4 (246 MB, 62.26%, 236 heads changed, −4 points on `paper_field`) loses accuracy. Grouped-channel palettization needs an iOS 18 deployment target and was not tried.
+
+## Packages
+
+| Bucket | Tokens | Heads × labels | Fast Decisions heads that fit | `fp16` | `w8` | `lut6` |
+| --- | ---: | ---: | ---: | --- | --- | --- |
+| `L128_H4_K8` | 128 | 4 × 8 | 10.7% | 8.2 ms | 10.1 ms | 9.9 ms |
+| `L256_H4_K32` | 256 | 4 × 32 | 95.0% | 14.7 ms | 20.6 ms | 17.6 ms |
+| `L512_H4_K32` | 512 | 4 × 32 | 100% | 44.0 ms | 42.9 ms | 62.9 ms |
+
+Latency is the p50 of 100 Python `predict` calls for a three-head request with compute units `ALL` on an M5 Pro with macOS 27.0, after 10 warmups (`benchmark-latency.py`). This is a local measurement for one request shape, not a device-wide benchmark; p95 values are in `reports/latency.json`. W8 and LUT6 are weight-only compression: they shrink the download and do not make inference faster. `CPU_AND_NE` is much slower than `ALL` for this graph (689 ms versus 14.7 ms p50 for `fp16` L256), so use `ALL`.
+
+On the first 20 bucket-fitting rows per domain, the L128 and L256 packages agree with the native predictions: `fp16` 173/173 and 345/345, `w8` 172/173 and 345/345, `lut6` 169/173 and 342/345 (`reports/verify-packages-L128-L256.json`).
+
+## Usage
+
+```bash
+uv sync
+uv run python runtime.py --model-dir . --precision lut6 --length 256 \
+  --text "My transfer is still pending and I used the wrong sort code. Can you stop it?" \
+  --tasks '{"intent": ["transfer_pending", "transfer_cancel", "beneficiary_add", "card_lost"], "urgency": ["low", "normal", "high"]}'
+```
+
+```python
+from runtime import CoreMLDecide
+
+model = CoreMLDecide(".", precision="w8", length=256)
+model.classify(
+    "Battery dies before lunch, but the keyboard and the screen are the best I have used.",
+    {"sentiment": ["positive", "negative", "mixed", "neutral"],
+     "aspects": {"labels": ["battery", "keyboard", "screen", "price"], "multi_label": True, "cls_threshold": 0.4}},
+)
+```
+
+The runtime loads only the tokenizer files and the Core ML package; it does not load PyTorch weights. Requests that exceed the bucket raise `ValueError`. When downloading with `huggingface_hub.snapshot_download`, pass `local_dir=` so the packages are materialized; Core ML compilation rejects the symlinked weight files in the default Hub cache.
+
+### Inputs and outputs
+
+| Name | Type | Shape | Meaning |
+| --- | --- | --- | --- |
+| `input_ids` | int32 | `[1, L]` | Schema and text tokens from the GLiNER2 span collator, padded |
+| `attention_mask` | int32 | `[1, L]` | 1 for real tokens |
+| `marker_indices` | int32 | `[1, 4, K]` | Token position of each `[L]` label marker, per head |
+| `marker_mask` | float32 | `[1, 4, K]` | 1 for used label slots |
+| `logits` | float32 | `[1, 4, K]` | Label logits; unused slots are −10⁴ |
+| `probabilities` | float32 | `[1, 4, K]` | Per-head softmax of `logits` |
+
+## Reproduce
+
+```bash
+uv sync
+uv run python convert-coreml.py --precision fp16 --length 512 --max-options 32
+uv run python compress-coreml.py --method linear \
+  --source build/gliner2_decide_classification_fp16_L512_H4_K32.mlpackage \
+  --output build/gliner2_decide_classification_w8_L512_H4_K32.mlpackage
+uv run python compress-coreml.py --method lut --bits 6 --granularity per_tensor \
+  --source build/gliner2_decide_classification_fp16_L512_H4_K32.mlpackage \
+  --output build/gliner2_decide_classification_lut6_L512_H4_K32.mlpackage
+uv run python score-fast-decisions.py --backend native --out build/fast-decisions-native.json
+uv run python score-fast-decisions.py --backend coreml \
+  --package build/gliner2_decide_classification_lut6_L512_H4_K32.mlpackage --out build/fast-decisions-lut6.json
+```
+
+`convert-coreml.py` checks the export wrapper against the native logits (4.8e-7) and the converted package against native `classify_text` on three examples. The Core ML deployment target is iOS 17 / macOS 14. Source file hashes are in `assets.lock.json`; conversion, scoring, agreement and latency reports are in `reports/`.

assets.lock.json

@@ -0,0 +1,32 @@
+{
+  "source": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "files": {
+    "config.json": {
+      "size": 594,
+      "sha256": "e748e5b80575471c91b3f0dd00f513ba58242544fcb7e1236e0021e61abd7673"
+    },
+    "encoder_config/config.json": {
+      "size": 920,
+      "sha256": "bd32f1484ba5a199f7a63df44df3814b839fffcf6e64478323c4689868ef6015"
+    },
+    "model.safetensors": {
+      "size": 1945828140,
+      "sha256": "40a5a23ff860dc3dff426cecd1048cacdd29c648c96db209dad818e9686dc997"
+    },
+    "special_tokens_map.json": {
+      "size": 2414,
+      "sha256": "84ea70143f533d7e99b393d87f20010887a9ac2cba955828ef313886e4e83f4f"
+    },
+    "tokenizer.json": {
+      "size": 8333952,
+      "sha256": "3ad87d9ffe669147063e70850927dd2da90249e2acc5c8527f1eb65df467bcc8"
+    },
+    "tokenizer_config.json": {
+      "size": 3356,
+      "sha256": "323199a4e946039410899f3779f2aa3eaef1500213c512727ad0f623d4f21309"
+    }
+  }
+}

benchmark-latency.py

@@ -0,0 +1,52 @@
+"""Median Core ML prediction latency for each package on one fixed three-head request."""
+import argparse
+import json
+import re
+import statistics
+import time
+from pathlib import Path
+
+import coremltools as ct
+from huggingface_hub import snapshot_download
+
+from convert_names import MODEL_ID, MODEL_REVISION
+from preprocessing import load_processor, prepare_decision
+
+TEXT = ("From: compliance@group.example\nSubject: Protocol update — action required today\n\n"
+        "Please confirm the new retention rule is applied before Friday's audit.")
+TASKS = {"intent": ["fyi", "request", "approval", "complaint", "newsletter", "security_alert"],
+         "urgency": ["low", "normal", "high", "critical"],
+         "route": ["support", "billing", "legal", "security", "finance", "archive"]}
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--warmup", type=int, default=10)
+    parser.add_argument("--calls", type=int, default=100)
+    parser.add_argument("--compute-units", choices=["all", "cpu_ne", "cpu_gpu", "cpu"], default="all")
+    parser.add_argument("--out", default="build/latency.json")
+    parser.add_argument("packages", nargs="+")
+    args = parser.parse_args()
+    processor = load_processor(snapshot_download(MODEL_ID, revision=MODEL_REVISION))
+    report = {}
+    for package in args.packages:
+        bucket = tuple(map(int, re.search(r"_L(\d+)_H(\d+)_K(\d+)", package).groups()))
+        arrays = prepare_decision(processor, TEXT, TASKS, *bucket)
+        units = {"all": ct.ComputeUnit.ALL, "cpu_ne": ct.ComputeUnit.CPU_AND_NE, "cpu_gpu": ct.ComputeUnit.CPU_AND_GPU,
+                 "cpu": ct.ComputeUnit.CPU_ONLY}[args.compute_units]
+        model = ct.models.MLModel(package, compute_units=units)
+        for _ in range(args.warmup):
+            model.predict(arrays)
+        times = []
+        for _ in range(args.calls):
+            start = time.perf_counter()
+            model.predict(arrays)
+            times.append((time.perf_counter() - start) * 1000)
+        report[Path(package).name] = {"p50_ms": statistics.median(times),
+                                      "p95_ms": sorted(times)[int(0.95 * (len(times) - 1))]}
+        print(Path(package).name, report[Path(package).name], flush=True)
+    Path(args.out).write_text(json.dumps(report, indent=2) + "\n")
+
+
+if __name__ == "__main__":
+    main()

compress-coreml.py

@@ -0,0 +1,86 @@
+"""Compress a verified GLiNER2.5-Decide classification Core ML model."""
+import argparse
+import json
+import time
+from pathlib import Path
+
+import coremltools as ct
+from coremltools.optimize.coreml import (
+    OpLinearQuantizerConfig,
+    OpPalettizerConfig,
+    OptimizationConfig,
+    get_weights_metadata,
+    linear_quantize_weights,
+    palettize_weights,
+)
+
+
+def package_bytes(path):
+    return sum(item.stat().st_size for item in path.rglob("*") if item.is_file())
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--source", required=True)
+    parser.add_argument("--output", required=True)
+    parser.add_argument("--bits", type=int, choices=[4, 6, 8], default=8)
+    parser.add_argument("--method", choices=["lut", "linear", "embedding"], default="lut")
+    parser.add_argument("--linear-mode", choices=["linear_symmetric", "linear"], default="linear_symmetric")
+    parser.add_argument("--granularity", choices=["per_tensor", "per_grouped_channel"], default="per_tensor")
+    parser.add_argument("--group-size", type=int, default=32)
+    args = parser.parse_args()
+    start = time.perf_counter()
+    model = ct.models.MLModel(args.source, skip_model_load=True)
+    selected_weights = None
+    if args.method in ("linear", "embedding"):
+        if args.bits != 8:
+            parser.error("linear quantization here supports only 8-bit weights")
+        quantizer = OpLinearQuantizerConfig(
+            mode=args.linear_mode, dtype="int8", granularity="per_channel", weight_threshold=2048,
+        )
+        if args.method == "embedding":
+            metadata = get_weights_metadata(model, weight_threshold=2048)
+            names = [
+                name for name, weight in metadata.items()
+                if "encoder_embeddings_word_embeddings" in name
+                and len(weight.val.shape) == 2
+                and any(op.op_type == "gather" for op in weight.child_ops)
+            ]
+            if len(names) != 1:
+                raise ValueError(f"expected one token embedding weight, found {names}")
+            config = OptimizationConfig(op_name_configs={names[0]: quantizer})
+            selected_weights = names
+        else:
+            config = OptimizationConfig(global_config=quantizer)
+        compressed = linear_quantize_weights(model, config=config)
+        variant = (
+            f"W8 {args.linear_mode} token embedding only"
+            if args.method == "embedding" else f"W8 {args.linear_mode} per-channel"
+        )
+    else:
+        config = OptimizationConfig(global_config=OpPalettizerConfig(
+            mode="kmeans", nbits=args.bits, granularity=args.granularity,
+            group_size=args.group_size,
+            enable_per_channel_scale=args.granularity == "per_grouped_channel",
+            num_kmeans_workers=4,
+        ))
+        compressed = palettize_weights(model, config=config)
+        variant = f"LUT{args.bits} {args.granularity}"
+    compressed.short_description = f"{model.short_description}; {variant} weights"
+    compressed.author = model.author
+    compressed.license = model.license
+    compressed.user_defined_metadata.update(model.user_defined_metadata)
+    compressed.user_defined_metadata["weight_compression"] = variant
+    compressed.save(args.output)
+    report = {"source": args.source, "output": args.output, "method": args.method,
+              "linear_mode": args.linear_mode if args.method in ("linear", "embedding") else None,
+              "granularity": "per_channel" if args.method in ("linear", "embedding") else args.granularity,
+              "selected_weights": selected_weights,
+              "source_bytes": package_bytes(Path(args.source)),
+              "output_bytes": package_bytes(Path(args.output)), "compression_seconds": time.perf_counter() - start}
+    Path(args.output).with_suffix(".json").write_text(json.dumps(report, indent=2) + "\n")
+    print(json.dumps(report, indent=2))
+
+
+if __name__ == "__main__":
+    main()

config.json

@@ -0,0 +1,24 @@
+{
+  "_attn_implementation_autoset": true,
+  "_name_or_path": "/home/urchadezaratiana/checkpoints/checkpoint-45500",
+  "architecture": "span",
+  "architecture_version": 1,
+  "architectures": [
+    "SpanExtractor"
+  ],
+  "attn_implementation": "sdpa",
+  "config_version": 3,
+  "counting_layer": "count_lstm",
+  "max_len": null,
+  "max_width": 8,
+  "model_name": "microsoft/deberta-v3-large",
+  "model_type": "extractor",
+  "span_head": {
+    "dropout": 0.1,
+    "max_width": 8,
+    "span_mode": "markerV0"
+  },
+  "token_pooling": "first",
+  "transformers_version": "4.48.1",
+  "use_moe": false
+}

convert-coreml.py

@@ -0,0 +1,168 @@
+"""Convert and verify the pinned GLiNER2.5-Decide multi-head classification path."""
+import argparse
+import json
+import math
+from pathlib import Path
+
+import coremltools as ct
+import numpy as np
+import torch
+from gliner2 import AutoExtractor
+from huggingface_hub import snapshot_download
+from transformers.models.deberta_v2 import modeling_deberta_v2
+
+from convert_names import MODEL_ID, MODEL_REVISION
+from export_model import GLiNER2DecideExport, coreml_safe_attention_forward
+from preprocessing import classification_schema, prepare_decision
+from runtime import decode, package_name
+
+EXAMPLES = [
+    ("My subscription renewed on April 15 for ¥5,400 after the service was already down. "
+     "Can I get that charge refunded?",
+     {"intent": ["order_status", "refund_request", "cancel_subscription", "update_payment", "login_problem",
+                 "shipping_delay", "bug_report", "speak_to_human"]}),
+    ("Battery dies before lunch, but the keyboard and the screen are the best I have used on a laptop.",
+     {"sentiment": ["positive", "negative", "mixed", "neutral"],
+      "aspects": {"labels": ["battery", "keyboard", "screen", "camera", "price", "support"],
+                  "multi_label": True, "cls_threshold": 0.4}}),
+    ("From: compliance@group.example\nSubject: Protocol update — action required today\n\n"
+     "Please confirm the new retention rule is applied before Friday's audit.",
+     {"intent": ["fyi", "request", "approval", "complaint", "newsletter", "security_alert"],
+      "urgency": ["low", "normal", "high", "critical"],
+      "route": ["support", "billing", "legal", "security", "finance", "archive"]}),
+]
+
+
+def native_logits(native, text, tasks):
+    """Per-head native logits via the upstream span collator, encoder and shared classifier."""
+    from gliner2.training.trainer import ExtractorCollator
+
+    collator = ExtractorCollator(native.processor, is_training=False, max_len=None, architecture="span")
+    batch = collator([(text, classification_schema(tasks).build())])
+    _, schema_embs = native._encode_batch(batch)
+    return [native.classifier(torch.stack(embs[1:])).squeeze(-1) for embs in schema_embs[0]]
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output-dir", default="build")
+    parser.add_argument("--length", type=int, default=128)
+    parser.add_argument("--max-heads", type=int, default=4)
+    parser.add_argument("--max-options", type=int, default=8)
+    parser.add_argument("--precision", choices=["fp16", "fp32"], default="fp16")
+    args = parser.parse_args()
+    torch.set_num_threads(4)
+    source = snapshot_download(
+        MODEL_ID, revision=MODEL_REVISION,
+        allow_patterns=[
+            "config.json", "encoder_config/*", "model.safetensors", "tokenizer.json", "tokenizer_config.json",
+            "special_tokens_map.json",
+        ],
+    )
+    native = AutoExtractor.from_pretrained(source, map_location="cpu").eval()
+    wrapper = GLiNER2DecideExport(native).eval()
+    bucket = (args.length, args.max_heads, args.max_options)
+    text, tasks = EXAMPLES[2]
+    arrays = prepare_decision(native.processor, text, tasks, *bucket)
+    tensors = tuple(torch.from_numpy(value) for value in arrays.values())
+
+    def wrapper_error():
+        with torch.no_grad():
+            expected = native_logits(native, text, tasks)
+            actual = wrapper(*tensors)[0][0]
+        return max(float((row - actual[h, : len(row)]).abs().max()) for h, row in enumerate(expected))
+
+    error = wrapper_error()
+    if error > 1e-4:
+        raise RuntimeError(f"Wrapper/native logit mismatch: {error}")
+    # The upstream scale is a constant for a fixed DeBERTa attention head width.
+    # Its traced int32 sqrt is rejected by Core ML; freeze the identical float32
+    # value while tracing, and restore the upstream implementation immediately.
+    original_scale = modeling_deberta_v2.scaled_size_sqrt
+    original_rpos = modeling_deberta_v2.build_rpos
+    original_attention = modeling_deberta_v2.DisentangledSelfAttention.forward
+
+    def static_scale(query_layer, scale_factor):
+        value = math.sqrt(float(query_layer.shape[-1] * scale_factor))
+        return torch.tensor(value, dtype=torch.float32, device=query_layer.device)
+
+    modeling_deberta_v2.scaled_size_sqrt = static_scale
+    # The encoder only uses self-attention: query and key sequence lengths are
+    # identical, so the scripted build_rpos returns relative_pos unchanged.
+    modeling_deberta_v2.build_rpos = lambda query, key, relative_pos, buckets, max_pos: relative_pos
+    modeling_deberta_v2.DisentangledSelfAttention.forward = coreml_safe_attention_forward
+    try:
+        frozen_error = wrapper_error()
+        if frozen_error > 1e-4:
+            raise RuntimeError(f"Frozen attention scale changed native logits: {frozen_error}")
+        with torch.no_grad():
+            traced = torch.jit.trace(wrapper, tensors, check_trace=False)
+    finally:
+        modeling_deberta_v2.scaled_size_sqrt = original_scale
+        modeling_deberta_v2.build_rpos = original_rpos
+        modeling_deberta_v2.DisentangledSelfAttention.forward = original_attention
+    grid = (1, args.max_heads, args.max_options)
+    converted = ct.convert(
+        traced, convert_to="mlprogram", minimum_deployment_target=ct.target.iOS17,
+        compute_precision=ct.precision.FLOAT16 if args.precision == "fp16" else ct.precision.FLOAT32,
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        inputs=[
+            ct.TensorType(name="input_ids", shape=(1, args.length), dtype=np.int32),
+            ct.TensorType(name="attention_mask", shape=(1, args.length), dtype=np.int32),
+            ct.TensorType(name="marker_indices", shape=grid, dtype=np.int32),
+            ct.TensorType(name="marker_mask", shape=grid, dtype=np.float32),
+        ],
+        outputs=[ct.TensorType(name="logits", dtype=np.float32), ct.TensorType(name="probabilities", dtype=np.float32)],
+    )
+    converted.short_description = "GLiNER2.5-Decide multi-head schema classification path"
+    converted.author = "Fastino (original); Fluid Inference (Core ML conversion)"
+    converted.license = "Apache-2.0"
+    converted.user_defined_metadata.update({
+        "source_model": MODEL_ID, "source_revision": MODEL_REVISION,
+        "scope": "classification only; span and count heads not exported",
+        "length": str(args.length), "max_heads": str(args.max_heads), "max_options": str(args.max_options),
+    })
+    out = Path(args.output_dir)
+    out.mkdir(parents=True, exist_ok=True)
+    package = out / package_name(args.precision, *bucket)
+    converted.save(str(package))
+    runtime = ct.models.MLModel(str(package), compute_units=ct.ComputeUnit.ALL)
+    cases = []
+    for text, tasks in EXAMPLES:
+        native_output = native.classify_text(text, tasks, include_confidence=True)
+        logits = np.asarray(runtime.predict(prepare_decision(native.processor, text, tasks, *bucket))["logits"])[0]
+        coreml_output = decode(tasks, logits)
+        if json.dumps(labels_only(native_output)) != json.dumps(labels_only(coreml_output)):
+            raise RuntimeError(f"Core ML/native mismatch: {coreml_output} != {native_output}")
+        cases.append({"text": text, "native": native_output, "coreml": coreml_output,
+                      "max_confidence_error": confidence_error(native_output, coreml_output)})
+    report = {
+        "source_model": MODEL_ID, "source_revision": MODEL_REVISION, "package": str(package),
+        "package_bytes": sum(f.stat().st_size for f in package.rglob("*") if f.is_file()),
+        "native_total_parameters": sum(p.numel() for p in native.parameters()),
+        "exported_parameters": sum(p.numel() for p in wrapper.parameters()),
+        "wrapper_max_logit_error": error, "coremltools": ct.__version__, "torch": torch.__version__,
+        "cases": cases,
+    }
+    (out / f"conversion-{package.stem}.json").write_text(json.dumps(report, indent=2, ensure_ascii=False) + "\n")
+    print(json.dumps(report, indent=2, ensure_ascii=False))
+
+
+def _entries(value):
+    return value if isinstance(value, list) else [value]
+
+
+def labels_only(result: dict) -> dict:
+    return {task: sorted(entry["label"] for entry in _entries(value)) for task, value in result.items()}
+
+
+def confidence_error(native: dict, coreml: dict) -> float:
+    errors = []
+    for task, value in native.items():
+        predicted = {entry["label"]: entry["confidence"] for entry in _entries(coreml[task])}
+        errors += [abs(entry["confidence"] - predicted[entry["label"]]) for entry in _entries(value)]
+    return max(errors)
+
+
+if __name__ == "__main__":
+    main()

convert_names.py

@@ -0,0 +1,3 @@
+"""Pinned source checkpoint shared by the conversion, verification and scoring scripts."""
+MODEL_ID = "fastino/GLiNER2.5-Decide"
+MODEL_REVISION = "65624f1a0265b3f612bae66a2685a06b94a68a9d"

encoder_config/config.json

@@ -0,0 +1,34 @@
+{
+  "_attn_implementation_autoset": true,
+  "_name_or_path": "/home/urchadezaratiana/checkpoints/checkpoint-45500/encoder_config/config.json",
+  "attention_probs_dropout_prob": 0.1,
+  "dtype": "float32",
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-07,
+  "legacy": true,
+  "max_position_embeddings": 512,
+  "max_relative_positions": -1,
+  "model_type": "deberta-v2",
+  "norm_rel_ebd": "layer_norm",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "pad_token_id": 0,
+  "pooler_dropout": 0,
+  "pooler_hidden_act": "gelu",
+  "pooler_hidden_size": 1024,
+  "pos_att_type": [
+    "p2c",
+    "c2p"
+  ],
+  "position_biased_input": false,
+  "position_buckets": 256,
+  "relative_attention": true,
+  "share_att_key": true,
+  "transformers_version": "4.48.1",
+  "type_vocab_size": 0,
+  "vocab_size": 128011
+}

export_model.py

@@ -0,0 +1,54 @@
+"""Native GLiNER2.5-Decide multi-head classification path with explicit marker routing."""
+import torch
+from torch import nn
+from transformers.models.deberta_v2 import modeling_deberta_v2
+
+
+def coreml_safe_attention_forward(
+    self, hidden_states, attention_mask, output_attentions=False,
+    query_states=None, relative_pos=None, rel_embeddings=None,
+):
+    """Native DeBERTa attention with a finite mask sentinel for FP16 Core ML."""
+    if query_states is None:
+        query_states = hidden_states
+    query = self.transpose_for_scores(self.query_proj(query_states), self.num_attention_heads)
+    key = self.transpose_for_scores(self.key_proj(hidden_states), self.num_attention_heads)
+    value = self.transpose_for_scores(self.value_proj(hidden_states), self.num_attention_heads)
+    factor = 1 + int("c2p" in self.pos_att_type) + int("p2c" in self.pos_att_type)
+    scale = modeling_deberta_v2.scaled_size_sqrt(query, factor)
+    scores = torch.bmm(query, key.transpose(-1, -2) / scale.to(dtype=query.dtype))
+    if self.relative_attention:
+        relative = self.disentangled_attention_bias(
+            query, key, relative_pos, self.pos_dropout(rel_embeddings), factor
+        )
+        scores = scores + relative
+    scores = scores.view(-1, self.num_attention_heads, scores.size(-2), scores.size(-1))
+    scores = scores.masked_fill(~attention_mask.bool(), -1e4)
+    probabilities = self.dropout(torch.softmax(scores, dim=-1))
+    context = torch.bmm(probabilities.view(-1, probabilities.size(-2), probabilities.size(-1)), value)
+    context = context.view(-1, self.num_attention_heads, context.size(-2), context.size(-1))
+    context = context.permute(0, 2, 1, 3).contiguous()
+    context = context.view(context.size()[:-2] + (-1,))
+    return (context, probabilities) if output_attentions else (context, None)
+
+
+class GLiNER2DecideExport(nn.Module):
+    """Encoder plus the shared label classifier, scored for every (head, label) marker at once.
+
+    Logits are returned per head so the host can apply softmax (single-label) or sigmoid
+    (multi-label) exactly as the native runtime does. Probabilities are the per-head softmax.
+    """
+
+    def __init__(self, native: nn.Module):
+        super().__init__()
+        self.encoder = native.encoder
+        self.classifier = native.classifier
+
+    def forward(self, input_ids, attention_mask, marker_indices, marker_mask):
+        hidden = self.encoder(input_ids=input_ids.long(), attention_mask=attention_mask.long()).last_hidden_state
+        heads, options = marker_indices.shape[1], marker_indices.shape[2]
+        flat = marker_indices.long().reshape(1, heads * options, 1).expand(-1, -1, hidden.shape[-1])
+        states = hidden.gather(1, flat)
+        logits = self.classifier(states).reshape(1, heads, options)
+        logits = torch.where(marker_mask > 0.5, logits, torch.full_like(logits, -1e4))
+        return logits, torch.softmax(logits, dim=-1)

gliner2_decide_classification_fp16_L128_H4_K8.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b61f4548bb86b1b61153b3a5d3a22637f54bc19493c6864e030493a2fbac88e
+size 1446151

gliner2_decide_classification_fp16_L128_H4_K8.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54501158f56baf0ebe295d99ac251eb0204cea2594881062b731bef9154032e2
+size 921351616

gliner2_decide_classification_fp16_L128_H4_K8.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "7194425C-FE01-46D6-87D7-5F4E30A614E1": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "7E5E91C3-28CD-4FE6-BA8B-03B642A946C5": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "7194425C-FE01-46D6-87D7-5F4E30A614E1"
+}

gliner2_decide_classification_fp16_L256_H4_K32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:626d74aac1448e0666e5a80d79302139dfd84d5ca80c44ec213c1c6a5497d823
+size 4331797

gliner2_decide_classification_fp16_L256_H4_K32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54501158f56baf0ebe295d99ac251eb0204cea2594881062b731bef9154032e2
+size 921351616

gliner2_decide_classification_fp16_L256_H4_K32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0F28C467-EA06-471F-A536-2845C1EA04D2": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "8A724345-8447-41EA-846B-8CF427172126": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "8A724345-8447-41EA-846B-8CF427172126"
+}

gliner2_decide_classification_fp16_L512_H4_K32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b215d623b792e4a107fc0a060c5d0ef921283f79ad45a1d275042419f2858564
+size 14821667

gliner2_decide_classification_fp16_L512_H4_K32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54501158f56baf0ebe295d99ac251eb0204cea2594881062b731bef9154032e2
+size 921351616

gliner2_decide_classification_fp16_L512_H4_K32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "2659AFFA-867C-410F-8AE6-0C4B3FC40276": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "39168FBF-EED8-4E6F-8E39-C3EB7295282B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "2659AFFA-867C-410F-8AE6-0C4B3FC40276"
+}

gliner2_decide_classification_lut6_L128_H4_K8.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f570cf44266c1707e8cdeafebba8084003c6399fab7a6b1961cb650e602733ab
+size 1475563

gliner2_decide_classification_lut6_L128_H4_K8.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:042712feeb1b1bc105603ec82b634eb05b1be58eeef878d840112b39bc8c8a93
+size 345968704

gliner2_decide_classification_lut6_L128_H4_K8.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0DA7B173-7F8F-44B1-80B4-525B4613CCC3": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "8533DAB7-36CE-4FCA-9EB3-D60A75900CDF": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "8533DAB7-36CE-4FCA-9EB3-D60A75900CDF"
+}

gliner2_decide_classification_lut6_L256_H4_K32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cdcb3bd54a8786aa05cdd8dbb1f34603ce3b0e64403d5973a3b971fb3ea036c
+size 4361209

gliner2_decide_classification_lut6_L256_H4_K32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:042712feeb1b1bc105603ec82b634eb05b1be58eeef878d840112b39bc8c8a93
+size 345968704

gliner2_decide_classification_lut6_L256_H4_K32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "BA61F74E-CA59-40B3-AADF-349CF3AED909": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "D5E517C8-12C8-4580-BFDC-5DE8AEFA3B99": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "BA61F74E-CA59-40B3-AADF-349CF3AED909"
+}

gliner2_decide_classification_lut6_L512_H4_K32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ace980c0a72efd57b3654772a542a0c4fbb6c9b669a9d1fa759eae970d4b98b
+size 14851079

gliner2_decide_classification_lut6_L512_H4_K32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:042712feeb1b1bc105603ec82b634eb05b1be58eeef878d840112b39bc8c8a93
+size 345968704

gliner2_decide_classification_lut6_L512_H4_K32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8EA312DA-C4A4-4D24-AF9E-6801A601CD31": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "D0A89611-DC9B-492B-ABFC-59000B9C7DD7": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "8EA312DA-C4A4-4D24-AF9E-6801A601CD31"
+}

gliner2_decide_classification_w8_L128_H4_K8.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9aaa4d0481ebea87aa200c9a5a6e6f2e091d802aadb54de2077f1e7dc734fcd0
+size 1489740

gliner2_decide_classification_w8_L128_H4_K8.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b509230b2fa7889c97086cc4c8da34adb5acab96feba286514a72dbdc681666
+size 461915648

gliner2_decide_classification_w8_L128_H4_K8.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "3E810636-76F0-484E-A35D-0C7E6A6C6C1A": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "ED9FF073-F8B5-4635-9513-F227567098FD": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "ED9FF073-F8B5-4635-9513-F227567098FD"
+}

gliner2_decide_classification_w8_L256_H4_K32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:999189808b876e9de8ebddbc33609881a84401a2dda9a64ecd82fdeb8722f95b
+size 4375386

gliner2_decide_classification_w8_L256_H4_K32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b509230b2fa7889c97086cc4c8da34adb5acab96feba286514a72dbdc681666
+size 461915648

gliner2_decide_classification_w8_L256_H4_K32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "37FC6834-7FAB-400D-A13F-DCBA6347753E": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "40CA9F0D-ADE7-424B-A117-8A5AFC0F0995": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "37FC6834-7FAB-400D-A13F-DCBA6347753E"
+}

gliner2_decide_classification_w8_L512_H4_K32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:400df7ec9300d4ccd0cd01e369df98ff45b24ae06a93774a6cd8591720220b39
+size 14865256

gliner2_decide_classification_w8_L512_H4_K32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b509230b2fa7889c97086cc4c8da34adb5acab96feba286514a72dbdc681666
+size 461915648

gliner2_decide_classification_w8_L512_H4_K32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "468F166A-BA65-4AE3-8D30-BDFC6296FD38": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "A069AFC1-E62E-49D5-AB90-C7DCCB259A61": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "468F166A-BA65-4AE3-8D30-BDFC6296FD38"
+}

preprocessing.py

@@ -0,0 +1,67 @@
+"""Native GLiNER2 span-architecture schema preprocessing for a fixed Core ML bucket."""
+import numpy as np
+from gliner2 import Schema
+from gliner2.models.base import load_extractor_tokenizer
+from gliner2.processor import SchemaTransformer
+from gliner2.training.trainer import ExtractorCollator
+
+
+def load_processor(tokenizer_dir: str):
+    """Load only the tokenizer and schema formatter needed by the Core ML model."""
+    return SchemaTransformer(tokenizer=load_extractor_tokenizer(tokenizer_dir), token_pooling="first")
+
+
+def classification_schema(tasks: dict) -> Schema:
+    """Same task-dict handling as native ``classify_text``."""
+    schema = Schema()
+    for name, config in tasks.items():
+        if isinstance(config, dict) and "labels" in config:
+            cfg = config.copy()
+            labels = cfg.pop("labels")
+            schema.classification(name, labels, **cfg)
+        else:
+            schema.classification(name, config)
+    return schema
+
+
+def task_labels(tasks: dict) -> dict[str, list[str]]:
+    result = {}
+    for name, config in tasks.items():
+        labels = config["labels"] if isinstance(config, dict) and "labels" in config else config
+        result[name] = list(labels.keys()) if isinstance(labels, dict) else list(labels)
+    return result
+
+
+def prepare_decision(processor, text: str, tasks: dict, length: int, max_heads: int, max_options: int):
+    """Tokenize ``tasks`` exactly as the native span collator does and pad into the bucket."""
+    if not 1 <= len(tasks) <= max_heads:
+        raise ValueError(f"Expected 1..{max_heads} decision heads, got {len(tasks)}")
+    labels = task_labels(tasks)
+    for name, values in labels.items():
+        if not 1 <= len(values) <= max_options:
+            raise ValueError(f"Head {name!r} has {len(values)} labels; bucket holds 1..{max_options}")
+    collator = ExtractorCollator(processor, is_training=False, max_len=None, architecture="span")
+    batch = collator([(text, classification_schema(tasks).build())])
+    ids = batch.input_ids.numpy()
+    if ids.shape[1] > length:
+        raise ValueError(f"Schema and text require {ids.shape[1]} subwords; bucket holds {length}")
+    groups = batch.schema_special_indices[0]
+    if len(groups) != len(labels):
+        raise ValueError("Schema head count does not match the requested tasks")
+    indices = np.zeros((1, max_heads, max_options), dtype=np.int32)
+    mask = np.zeros((1, max_heads, max_options), dtype=np.float32)
+    for head, (positions, values) in enumerate(zip(groups, labels.values())):
+        # positions[0] is the [P] prompt marker; the rest are one [L] marker per label.
+        markers = list(positions[1:])
+        if len(markers) != len(values):
+            raise ValueError("Label markers were truncated or merged")
+        indices[0, head, : len(markers)] = markers
+        mask[0, head, : len(markers)] = 1.0
+    attention = batch.attention_mask.numpy()
+    pad = processor.tokenizer.pad_token_id
+    return {
+        "input_ids": np.pad(ids, ((0, 0), (0, length - ids.shape[1])), constant_values=pad).astype(np.int32),
+        "attention_mask": np.pad(attention, ((0, 0), (0, length - attention.shape[1]))).astype(np.int32),
+        "marker_indices": indices,
+        "marker_mask": mask,
+    }

pyproject.toml

@@ -0,0 +1,31 @@
+[project]
+name = "gliner2-decide-coreml"
+version = "0.1.0"
+description = "Pinned GLiNER2.5-Decide Core ML classification export"
+requires-python = ">=3.12,<3.13"
+dependencies = [
+    "coremltools==9.0",
+    "gliner2[local]==2.0.0",
+    "huggingface-hub>=0.34,<1",
+    "numpy<2.3",
+    "protobuf>=5,<7",
+    "psutil>=7,<8",
+    "sentencepiece>=0.2,<0.3",
+    "torch==2.7.0",
+    "transformers==4.57.6",
+]
+
+[dependency-groups]
+dev = ["pytest>=8.4", "ruff>=0.13"]
+compression = ["scikit-learn==1.5.1"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+pythonpath = ["."]
+
+[tool.ruff]
+line-length = 120
+target-version = "py312"
+
+[tool.ruff.lint]
+select = ["E", "F", "I"]

reports/conversion-gliner2_decide_classification_fp16_L128_H4_K8.json

@@ -0,0 +1,105 @@
+{
+  "source_model": "fastino/GLiNER2.5-Decide",
+  "source_revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "package": "build/gliner2_decide_classification_fp16_L128_H4_K8.mlpackage",
+  "package_bytes": 922798384,
+  "native_total_parameters": 486444053,
+  "exported_parameters": 436022273,
+  "wrapper_max_logit_error": 4.76837158203125e-07,
+  "coremltools": "9.0",
+  "torch": "2.7.0",
+  "cases": [
+    {
+      "text": "My subscription renewed on April 15 for ¥5,400 after the service was already down. Can I get that charge refunded?",
+      "native": {
+        "intent": {
+          "label": "refund_request",
+          "confidence": 0.9982081651687622
+        }
+      },
+      "coreml": {
+        "intent": {
+          "label": "refund_request",
+          "confidence": 0.9982183726849228
+        }
+      },
+      "max_confidence_error": 1.0207516160543051e-05
+    },
+    {
+      "text": "Battery dies before lunch, but the keyboard and the screen are the best I have used on a laptop.",
+      "native": {
+        "sentiment": {
+          "label": "positive",
+          "confidence": 0.9994937181472778
+        },
+        "aspects": [
+          {
+            "label": "battery",
+            "confidence": 0.7610353231430054
+          },
+          {
+            "label": "keyboard",
+            "confidence": 0.9939987659454346
+          },
+          {
+            "label": "screen",
+            "confidence": 0.9846210479736328
+          }
+        ]
+      },
+      "coreml": {
+        "sentiment": {
+          "label": "positive",
+          "confidence": 0.9994971447974256
+        },
+        "aspects": [
+          {
+            "label": "battery",
+            "confidence": 0.7611836278339746
+          },
+          {
+            "label": "keyboard",
+            "confidence": 0.9940428462362503
+          },
+          {
+            "label": "screen",
+            "confidence": 0.9846346585586978
+          }
+        ]
+      },
+      "max_confidence_error": 0.0001483046909692698
+    },
+    {
+      "text": "From: compliance@group.example\nSubject: Protocol update — action required today\n\nPlease confirm the new retention rule is applied before Friday's audit.",
+      "native": {
+        "intent": {
+          "label": "request",
+          "confidence": 0.4523202180862427
+        },
+        "urgency": {
+          "label": "critical",
+          "confidence": 0.4121380150318146
+        },
+        "route": {
+          "label": "legal",
+          "confidence": 0.5795720219612122
+        }
+      },
+      "coreml": {
+        "intent": {
+          "label": "request",
+          "confidence": 0.45214515554103
+        },
+        "urgency": {
+          "label": "critical",
+          "confidence": 0.4120186418236959
+        },
+        "route": {
+          "label": "legal",
+          "confidence": 0.5799514106438567
+        }
+      },
+      "max_confidence_error": 0.0003793886826445325
+    }
+  ]
+}

reports/conversion-gliner2_decide_classification_fp16_L256_H4_K32.json

@@ -0,0 +1,105 @@
+{
+  "source_model": "fastino/GLiNER2.5-Decide",
+  "source_revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "package": "build/gliner2_decide_classification_fp16_L256_H4_K32.mlpackage",
+  "package_bytes": 925684030,
+  "native_total_parameters": 486444053,
+  "exported_parameters": 436022273,
+  "wrapper_max_logit_error": 4.76837158203125e-07,
+  "coremltools": "9.0",
+  "torch": "2.7.0",
+  "cases": [
+    {
+      "text": "My subscription renewed on April 15 for ¥5,400 after the service was already down. Can I get that charge refunded?",
+      "native": {
+        "intent": {
+          "label": "refund_request",
+          "confidence": 0.9982081651687622
+        }
+      },
+      "coreml": {
+        "intent": {
+          "label": "refund_request",
+          "confidence": 0.9982183726849228
+        }
+      },
+      "max_confidence_error": 1.0207516160543051e-05
+    },
+    {
+      "text": "Battery dies before lunch, but the keyboard and the screen are the best I have used on a laptop.",
+      "native": {
+        "sentiment": {
+          "label": "positive",
+          "confidence": 0.9994937181472778
+        },
+        "aspects": [
+          {
+            "label": "battery",
+            "confidence": 0.7610353231430054
+          },
+          {
+            "label": "keyboard",
+            "confidence": 0.9939987659454346
+          },
+          {
+            "label": "screen",
+            "confidence": 0.9846210479736328
+          }
+        ]
+      },
+      "coreml": {
+        "sentiment": {
+          "label": "positive",
+          "confidence": 0.9994971447974256
+        },
+        "aspects": [
+          {
+            "label": "battery",
+            "confidence": 0.7611836278339746
+          },
+          {
+            "label": "keyboard",
+            "confidence": 0.9940428462362503
+          },
+          {
+            "label": "screen",
+            "confidence": 0.9846346585586978
+          }
+        ]
+      },
+      "max_confidence_error": 0.0001483046909692698
+    },
+    {
+      "text": "From: compliance@group.example\nSubject: Protocol update — action required today\n\nPlease confirm the new retention rule is applied before Friday's audit.",
+      "native": {
+        "intent": {
+          "label": "request",
+          "confidence": 0.4523202180862427
+        },
+        "urgency": {
+          "label": "critical",
+          "confidence": 0.4121380150318146
+        },
+        "route": {
+          "label": "legal",
+          "confidence": 0.5795720219612122
+        }
+      },
+      "coreml": {
+        "intent": {
+          "label": "request",
+          "confidence": 0.45214515554103
+        },
+        "urgency": {
+          "label": "critical",
+          "confidence": 0.4120186418236959
+        },
+        "route": {
+          "label": "legal",
+          "confidence": 0.5799514106438567
+        }
+      },
+      "max_confidence_error": 0.0003793886826445325
+    }
+  ]
+}

reports/conversion-gliner2_decide_classification_fp16_L512_H4_K32.json

@@ -0,0 +1,105 @@
+{
+  "source_model": "fastino/GLiNER2.5-Decide",
+  "source_revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "package": "build/gliner2_decide_classification_fp16_L512_H4_K32.mlpackage",
+  "package_bytes": 936173900,
+  "native_total_parameters": 486444053,
+  "exported_parameters": 436022273,
+  "wrapper_max_logit_error": 4.76837158203125e-07,
+  "coremltools": "9.0",
+  "torch": "2.7.0",
+  "cases": [
+    {
+      "text": "My subscription renewed on April 15 for ¥5,400 after the service was already down. Can I get that charge refunded?",
+      "native": {
+        "intent": {
+          "label": "refund_request",
+          "confidence": 0.9982081651687622
+        }
+      },
+      "coreml": {
+        "intent": {
+          "label": "refund_request",
+          "confidence": 0.9982183726849228
+        }
+      },
+      "max_confidence_error": 1.0207516160543051e-05
+    },
+    {
+      "text": "Battery dies before lunch, but the keyboard and the screen are the best I have used on a laptop.",
+      "native": {
+        "sentiment": {
+          "label": "positive",
+          "confidence": 0.9994937181472778
+        },
+        "aspects": [
+          {
+            "label": "battery",
+            "confidence": 0.7610353231430054
+          },
+          {
+            "label": "keyboard",
+            "confidence": 0.9939987659454346
+          },
+          {
+            "label": "screen",
+            "confidence": 0.9846210479736328
+          }
+        ]
+      },
+      "coreml": {
+        "sentiment": {
+          "label": "positive",
+          "confidence": 0.9994971447974256
+        },
+        "aspects": [
+          {
+            "label": "battery",
+            "confidence": 0.7611836278339746
+          },
+          {
+            "label": "keyboard",
+            "confidence": 0.9940428462362503
+          },
+          {
+            "label": "screen",
+            "confidence": 0.9846346585586978
+          }
+        ]
+      },
+      "max_confidence_error": 0.0001483046909692698
+    },
+    {
+      "text": "From: compliance@group.example\nSubject: Protocol update — action required today\n\nPlease confirm the new retention rule is applied before Friday's audit.",
+      "native": {
+        "intent": {
+          "label": "request",
+          "confidence": 0.4523202180862427
+        },
+        "urgency": {
+          "label": "critical",
+          "confidence": 0.4121380150318146
+        },
+        "route": {
+          "label": "legal",
+          "confidence": 0.5795720219612122
+        }
+      },
+      "coreml": {
+        "intent": {
+          "label": "request",
+          "confidence": 0.45214515554103
+        },
+        "urgency": {
+          "label": "critical",
+          "confidence": 0.4120186418236959
+        },
+        "route": {
+          "label": "legal",
+          "confidence": 0.5799514106438567
+        }
+      },
+      "max_confidence_error": 0.0003793886826445325
+    }
+  ]
+}

reports/fast-decisions-coreml-embedding_w8-L512-all.json

@@ -0,0 +1,102 @@
+{
+  "model": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "split": "development (public, 100 rows per domain); published card scores use the held-out test split",
+  "backend": "coreml",
+  "package": "build/gliner2_decide_classification_embedding_w8_L512_H4_K32.mlpackage",
+  "compute_units": "all",
+  "average_accuracy": 0.6299509803921568,
+  "pooled_accuracy": 0.6141379310344828,
+  "model_call_p50_ms": 41.95125005207956,
+  "model_call_p95_ms": 45.35062506329268,
+  "seconds": 133.38757445802912,
+  "domains": {
+    "support_intent": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "support_topic": {
+      "heads": 100,
+      "hits": 44,
+      "accuracy": 0.44
+    },
+    "document_type": {
+      "heads": 100,
+      "hits": 79,
+      "accuracy": 0.79
+    },
+    "review_sentiment": {
+      "heads": 100,
+      "hits": 84,
+      "accuracy": 0.84
+    },
+    "agent_handoff": {
+      "heads": 100,
+      "hits": 77,
+      "accuracy": 0.77
+    },
+    "email_triage": {
+      "heads": 400,
+      "hits": 237,
+      "accuracy": 0.5925
+    },
+    "ticket_route": {
+      "heads": 300,
+      "hits": 145,
+      "accuracy": 0.48333333333333334
+    },
+    "product_feedback": {
+      "heads": 200,
+      "hits": 104,
+      "accuracy": 0.52
+    },
+    "banking_intent": {
+      "heads": 100,
+      "hits": 64,
+      "accuracy": 0.64
+    },
+    "clinic_request": {
+      "heads": 200,
+      "hits": 129,
+      "accuracy": 0.645
+    },
+    "travel_request": {
+      "heads": 100,
+      "hits": 63,
+      "accuracy": 0.63
+    },
+    "news_topic": {
+      "heads": 100,
+      "hits": 69,
+      "accuracy": 0.69
+    },
+    "paper_field": {
+      "heads": 100,
+      "hits": 47,
+      "accuracy": 0.47
+    },
+    "sports_recap": {
+      "heads": 300,
+      "hits": 226,
+      "accuracy": 0.7533333333333333
+    },
+    "restaurant_review": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    },
+    "benefits_request": {
+      "heads": 200,
+      "hits": 151,
+      "accuracy": 0.755
+    },
+    "screen_tags": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    }
+  }
+}

reports/fast-decisions-coreml-fp16-L512-all.json

@@ -0,0 +1,102 @@
+{
+  "model": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "split": "development (public, 100 rows per domain); published card scores use the held-out test split",
+  "backend": "coreml",
+  "package": "build/gliner2_decide_classification_fp16_L512_H4_K32.mlpackage",
+  "compute_units": "all",
+  "average_accuracy": 0.629264705882353,
+  "pooled_accuracy": 0.6137931034482759,
+  "model_call_p50_ms": 49.66187500394881,
+  "model_call_p95_ms": 64.96874999720603,
+  "seconds": 158.56965945789125,
+  "domains": {
+    "support_intent": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "support_topic": {
+      "heads": 100,
+      "hits": 44,
+      "accuracy": 0.44
+    },
+    "document_type": {
+      "heads": 100,
+      "hits": 79,
+      "accuracy": 0.79
+    },
+    "review_sentiment": {
+      "heads": 100,
+      "hits": 84,
+      "accuracy": 0.84
+    },
+    "agent_handoff": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "email_triage": {
+      "heads": 400,
+      "hits": 237,
+      "accuracy": 0.5925
+    },
+    "ticket_route": {
+      "heads": 300,
+      "hits": 145,
+      "accuracy": 0.48333333333333334
+    },
+    "product_feedback": {
+      "heads": 200,
+      "hits": 104,
+      "accuracy": 0.52
+    },
+    "banking_intent": {
+      "heads": 100,
+      "hits": 64,
+      "accuracy": 0.64
+    },
+    "clinic_request": {
+      "heads": 200,
+      "hits": 129,
+      "accuracy": 0.645
+    },
+    "travel_request": {
+      "heads": 100,
+      "hits": 63,
+      "accuracy": 0.63
+    },
+    "news_topic": {
+      "heads": 100,
+      "hits": 69,
+      "accuracy": 0.69
+    },
+    "paper_field": {
+      "heads": 100,
+      "hits": 47,
+      "accuracy": 0.47
+    },
+    "sports_recap": {
+      "heads": 300,
+      "hits": 227,
+      "accuracy": 0.7566666666666667
+    },
+    "restaurant_review": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    },
+    "benefits_request": {
+      "heads": 200,
+      "hits": 150,
+      "accuracy": 0.75
+    },
+    "screen_tags": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    }
+  }
+}

reports/fast-decisions-coreml-lut4-L512-all.json

@@ -0,0 +1,102 @@
+{
+  "model": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "split": "development (public, 100 rows per domain); published card scores use the held-out test split",
+  "backend": "coreml",
+  "package": "build/gliner2_decide_classification_lut4_L512_H4_K32.mlpackage",
+  "compute_units": "all",
+  "average_accuracy": 0.6226470588235294,
+  "pooled_accuracy": 0.606551724137931,
+  "model_call_p50_ms": 43.17912494298071,
+  "model_call_p95_ms": 48.890333040617406,
+  "seconds": 132.74362174991984,
+  "domains": {
+    "support_intent": {
+      "heads": 100,
+      "hits": 75,
+      "accuracy": 0.75
+    },
+    "support_topic": {
+      "heads": 100,
+      "hits": 46,
+      "accuracy": 0.46
+    },
+    "document_type": {
+      "heads": 100,
+      "hits": 78,
+      "accuracy": 0.78
+    },
+    "review_sentiment": {
+      "heads": 100,
+      "hits": 85,
+      "accuracy": 0.85
+    },
+    "agent_handoff": {
+      "heads": 100,
+      "hits": 73,
+      "accuracy": 0.73
+    },
+    "email_triage": {
+      "heads": 400,
+      "hits": 234,
+      "accuracy": 0.585
+    },
+    "ticket_route": {
+      "heads": 300,
+      "hits": 143,
+      "accuracy": 0.4766666666666667
+    },
+    "product_feedback": {
+      "heads": 200,
+      "hits": 102,
+      "accuracy": 0.51
+    },
+    "banking_intent": {
+      "heads": 100,
+      "hits": 63,
+      "accuracy": 0.63
+    },
+    "clinic_request": {
+      "heads": 200,
+      "hits": 129,
+      "accuracy": 0.645
+    },
+    "travel_request": {
+      "heads": 100,
+      "hits": 64,
+      "accuracy": 0.64
+    },
+    "news_topic": {
+      "heads": 100,
+      "hits": 70,
+      "accuracy": 0.7
+    },
+    "paper_field": {
+      "heads": 100,
+      "hits": 43,
+      "accuracy": 0.43
+    },
+    "sports_recap": {
+      "heads": 300,
+      "hits": 223,
+      "accuracy": 0.7433333333333333
+    },
+    "restaurant_review": {
+      "heads": 200,
+      "hits": 90,
+      "accuracy": 0.45
+    },
+    "benefits_request": {
+      "heads": 200,
+      "hits": 151,
+      "accuracy": 0.755
+    },
+    "screen_tags": {
+      "heads": 200,
+      "hits": 90,
+      "accuracy": 0.45
+    }
+  }
+}

reports/fast-decisions-coreml-lut6-L512-all.json

@@ -0,0 +1,102 @@
+{
+  "model": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "split": "development (public, 100 rows per domain); published card scores use the held-out test split",
+  "backend": "coreml",
+  "package": "build/gliner2_decide_classification_lut6_L512_H4_K32.mlpackage",
+  "compute_units": "all",
+  "average_accuracy": 0.6302941176470588,
+  "pooled_accuracy": 0.6151724137931035,
+  "model_call_p50_ms": 41.30370810162276,
+  "model_call_p95_ms": 45.43866601306945,
+  "seconds": 125.51691958296578,
+  "domains": {
+    "support_intent": {
+      "heads": 100,
+      "hits": 75,
+      "accuracy": 0.75
+    },
+    "support_topic": {
+      "heads": 100,
+      "hits": 45,
+      "accuracy": 0.45
+    },
+    "document_type": {
+      "heads": 100,
+      "hits": 79,
+      "accuracy": 0.79
+    },
+    "review_sentiment": {
+      "heads": 100,
+      "hits": 84,
+      "accuracy": 0.84
+    },
+    "agent_handoff": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "email_triage": {
+      "heads": 400,
+      "hits": 234,
+      "accuracy": 0.585
+    },
+    "ticket_route": {
+      "heads": 300,
+      "hits": 149,
+      "accuracy": 0.49666666666666665
+    },
+    "product_feedback": {
+      "heads": 200,
+      "hits": 103,
+      "accuracy": 0.515
+    },
+    "banking_intent": {
+      "heads": 100,
+      "hits": 64,
+      "accuracy": 0.64
+    },
+    "clinic_request": {
+      "heads": 200,
+      "hits": 129,
+      "accuracy": 0.645
+    },
+    "travel_request": {
+      "heads": 100,
+      "hits": 63,
+      "accuracy": 0.63
+    },
+    "news_topic": {
+      "heads": 100,
+      "hits": 69,
+      "accuracy": 0.69
+    },
+    "paper_field": {
+      "heads": 100,
+      "hits": 47,
+      "accuracy": 0.47
+    },
+    "sports_recap": {
+      "heads": 300,
+      "hits": 229,
+      "accuracy": 0.7633333333333333
+    },
+    "restaurant_review": {
+      "heads": 200,
+      "hits": 94,
+      "accuracy": 0.47
+    },
+    "benefits_request": {
+      "heads": 200,
+      "hits": 152,
+      "accuracy": 0.76
+    },
+    "screen_tags": {
+      "heads": 200,
+      "hits": 92,
+      "accuracy": 0.46
+    }
+  }
+}

reports/fast-decisions-coreml-w8-L512-all.json

@@ -0,0 +1,102 @@
+{
+  "model": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "split": "development (public, 100 rows per domain); published card scores use the held-out test split",
+  "backend": "coreml",
+  "package": "build/gliner2_decide_classification_w8_L512_H4_K32.mlpackage",
+  "compute_units": "all",
+  "average_accuracy": 0.6301470588235294,
+  "pooled_accuracy": 0.6151724137931035,
+  "model_call_p50_ms": 46.61016596946865,
+  "model_call_p95_ms": 52.74316598661244,
+  "seconds": 150.09406595793553,
+  "domains": {
+    "support_intent": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "support_topic": {
+      "heads": 100,
+      "hits": 44,
+      "accuracy": 0.44
+    },
+    "document_type": {
+      "heads": 100,
+      "hits": 79,
+      "accuracy": 0.79
+    },
+    "review_sentiment": {
+      "heads": 100,
+      "hits": 84,
+      "accuracy": 0.84
+    },
+    "agent_handoff": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "email_triage": {
+      "heads": 400,
+      "hits": 239,
+      "accuracy": 0.5975
+    },
+    "ticket_route": {
+      "heads": 300,
+      "hits": 146,
+      "accuracy": 0.4866666666666667
+    },
+    "product_feedback": {
+      "heads": 200,
+      "hits": 104,
+      "accuracy": 0.52
+    },
+    "banking_intent": {
+      "heads": 100,
+      "hits": 64,
+      "accuracy": 0.64
+    },
+    "clinic_request": {
+      "heads": 200,
+      "hits": 129,
+      "accuracy": 0.645
+    },
+    "travel_request": {
+      "heads": 100,
+      "hits": 63,
+      "accuracy": 0.63
+    },
+    "news_topic": {
+      "heads": 100,
+      "hits": 69,
+      "accuracy": 0.69
+    },
+    "paper_field": {
+      "heads": 100,
+      "hits": 47,
+      "accuracy": 0.47
+    },
+    "sports_recap": {
+      "heads": 300,
+      "hits": 226,
+      "accuracy": 0.7533333333333333
+    },
+    "restaurant_review": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    },
+    "benefits_request": {
+      "heads": 200,
+      "hits": 151,
+      "accuracy": 0.755
+    },
+    "screen_tags": {
+      "heads": 200,
+      "hits": 94,
+      "accuracy": 0.47
+    }
+  }
+}

reports/fast-decisions-native.json

@@ -0,0 +1,102 @@
+{
+  "model": "fastino/GLiNER2.5-Decide",
+  "revision": "65624f1a0265b3f612bae66a2685a06b94a68a9d",
+  "dataset": "fastino/fast-decisions",
+  "dataset_revision": "1a33070cabf94ce2e29105482dd2ef6c157ad7f2",
+  "split": "development (public, 100 rows per domain); published card scores use the held-out test split",
+  "backend": "native",
+  "package": null,
+  "compute_units": null,
+  "average_accuracy": 0.629264705882353,
+  "pooled_accuracy": 0.6137931034482759,
+  "model_call_p50_ms": null,
+  "model_call_p95_ms": null,
+  "seconds": 794.7205639999593,
+  "domains": {
+    "support_intent": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "support_topic": {
+      "heads": 100,
+      "hits": 44,
+      "accuracy": 0.44
+    },
+    "document_type": {
+      "heads": 100,
+      "hits": 79,
+      "accuracy": 0.79
+    },
+    "review_sentiment": {
+      "heads": 100,
+      "hits": 84,
+      "accuracy": 0.84
+    },
+    "agent_handoff": {
+      "heads": 100,
+      "hits": 76,
+      "accuracy": 0.76
+    },
+    "email_triage": {
+      "heads": 400,
+      "hits": 237,
+      "accuracy": 0.5925
+    },
+    "ticket_route": {
+      "heads": 300,
+      "hits": 145,
+      "accuracy": 0.48333333333333334
+    },
+    "product_feedback": {
+      "heads": 200,
+      "hits": 104,
+      "accuracy": 0.52
+    },
+    "banking_intent": {
+      "heads": 100,
+      "hits": 64,
+      "accuracy": 0.64
+    },
+    "clinic_request": {
+      "heads": 200,
+      "hits": 129,
+      "accuracy": 0.645
+    },
+    "travel_request": {
+      "heads": 100,
+      "hits": 63,
+      "accuracy": 0.63
+    },
+    "news_topic": {
+      "heads": 100,
+      "hits": 69,
+      "accuracy": 0.69
+    },
+    "paper_field": {
+      "heads": 100,
+      "hits": 47,
+      "accuracy": 0.47
+    },
+    "sports_recap": {
+      "heads": 300,
+      "hits": 227,
+      "accuracy": 0.7566666666666667
+    },
+    "restaurant_review": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    },
+    "benefits_request": {
+      "heads": 200,
+      "hits": 150,
+      "accuracy": 0.75
+    },
+    "screen_tags": {
+      "heads": 200,
+      "hits": 93,
+      "accuracy": 0.465
+    }
+  }
+}

reports/latency.json

@@ -0,0 +1,47 @@
+{
+  "machine": "Apple M5 Pro, macOS 27.0, compute units ALL",
+  "request": "three-head email triage request (benchmark-latency.py), 10 warmups, 100 calls, Python coremltools predict",
+  "gliner2_decide_classification_fp16_L128_H4_K8.mlpackage": {
+    "p50_ms": 8.18,
+    "p95_ms": 10.33
+  },
+  "gliner2_decide_classification_w8_L128_H4_K8.mlpackage": {
+    "p50_ms": 10.09,
+    "p95_ms": 12.98
+  },
+  "gliner2_decide_classification_lut6_L128_H4_K8.mlpackage": {
+    "p50_ms": 9.89,
+    "p95_ms": 14.45
+  },
+  "gliner2_decide_classification_fp16_L256_H4_K32.mlpackage": {
+    "p50_ms": 14.71,
+    "p95_ms": 16.69
+  },
+  "gliner2_decide_classification_w8_L256_H4_K32.mlpackage": {
+    "p50_ms": 20.58,
+    "p95_ms": 38.79
+  },
+  "gliner2_decide_classification_lut6_L256_H4_K32.mlpackage": {
+    "p50_ms": 17.58,
+    "p95_ms": 24.61
+  },
+  "gliner2_decide_classification_fp16_L512_H4_K32.mlpackage": {
+    "p50_ms": 44.02,
+    "p95_ms": 98.99
+  },
+  "gliner2_decide_classification_w8_L512_H4_K32.mlpackage": {
+    "p50_ms": 42.93,
+    "p95_ms": 66.93
+  },
+  "gliner2_decide_classification_lut6_L512_H4_K32.mlpackage": {
+    "p50_ms": 62.95,
+    "p95_ms": 71.11
+  },
+  "cpu_and_ne": {
+    "gliner2_decide_classification_fp16_L256_H4_K32.mlpackage": {
+      "p50_ms": 689.15,
+      "p95_ms": 1266.79,
+      "calls": 30
+    }
+  }
+}

reports/verify-packages-L128-L256.json

@@ -0,0 +1,26 @@
+{
+  "gliner2_decide_classification_fp16_L128_H4_K8.mlpackage": {
+    "checked_heads": 173,
+    "agree_with_native": 173
+  },
+  "gliner2_decide_classification_w8_L128_H4_K8.mlpackage": {
+    "checked_heads": 173,
+    "agree_with_native": 172
+  },
+  "gliner2_decide_classification_lut6_L128_H4_K8.mlpackage": {
+    "checked_heads": 173,
+    "agree_with_native": 169
+  },
+  "gliner2_decide_classification_fp16_L256_H4_K32.mlpackage": {
+    "checked_heads": 345,
+    "agree_with_native": 345
+  },
+  "gliner2_decide_classification_w8_L256_H4_K32.mlpackage": {
+    "checked_heads": 345,
+    "agree_with_native": 345
+  },
+  "gliner2_decide_classification_lut6_L256_H4_K32.mlpackage": {
+    "checked_heads": 345,
+    "agree_with_native": 342
+  }
+}