sb-decision 0.2.0__tar.gz

sb_decision-0.2.0/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 2026 Sarathi Balakrishnan
+
+   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.

sb_decision-0.2.0/NOTICE

@@ -0,0 +1,12 @@
+SB Decision Model (Browser Decision Model)
+Copyright 2026 Sarathi Balakrishnan
+
+Licensed under the Apache License, Version 2.0. Redistributions must keep this NOTICE.
+
+Built on:
+- ModernBERT (Answer.AI / LightOn), Apache-2.0, the encoder backbone.
+- GLiClass (Knowledgator), Apache-2.0, the label-token scoring recipe and the pretrained checkpoint the
+  fine-tune started from.
+
+The decision framing, option-set design, synthetic data recipe, trained-in abstention, calibration, ONNX export
+recipe, browser runtime and documentation are original work by Sarathi Balakrishnan.

sb_decision-0.2.0/PKG-INFO

@@ -0,0 +1,105 @@
+Metadata-Version: 2.4
+Name: sb-decision
+Version: 0.2.0
+Summary: SB Decision Model: score your own options for any text in one pass, with a trained-in ask. By Sarathi Balakrishnan.
+Author: Sarathi Balakrishnan
+License: Apache-2.0
+Project-URL: Model, https://huggingface.co/sarathi-balakrishnan/browser-decision-model-base
+Project-URL: Author, https://huggingface.co/sarathi-balakrishnan
+Keywords: decision,routing,triage,classification,zero-shot,guardrail
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: torch>=2.2
+Requires-Dist: transformers>=4.48
+Requires-Dist: huggingface_hub>=0.24
+Dynamic: license-file
+
+# sb-decision
+
+The SB Decision Model, by Sarathi Balakrishnan. Give it a text and your own list of options, each written as
+`name: when to choose it`, and it returns a calibrated probability per option in one pass. Nothing is retrained per
+use case: the option definitions are the policy. A reserved `ask` option is trained to win when the deciding fact is
+missing from the text, so irreversible actions go to a person instead of being guessed.
+
+Not an LLM. It never generates text. It is the router, guardrail, or triage layer in front of one.
+
+## Install
+
+```
+pip install sb-decision
+```
+
+## Use
+
+```python
+from sb_decision import SBDecision
+
+m = SBDecision.load()  # sarathi-balakrishnan/browser-decision-model-base
+m.decide(
+    "Package arrived broken, second time this month. I want a manager.",
+    [
+        "refund: customer wants money back for a damaged, wrong, or missing item",
+        "replace: customer wants the product again",
+        "escalate: angry, threatens to leave, or asks for a manager",
+    ],
+)
+# {'decision': 'escalate', 'band': 'auto', 'scores': {'escalate': 0.90, 'ask': 0.09, 'refund': 0.01, 'replace': 0.01}}
+```
+
+`band` is `auto` when the top score is at or above `auto=0.7`, `review` at or above `review=0.4`, otherwise
+`abstain`. If `ask` wins it is always `abstain`. Both thresholds are arguments; they are statements about the error
+rate you accept, not model constants.
+
+Independent tags, when several options can apply at once:
+
+```python
+m.decide(
+    "Headache since morning and now my chest feels tight.",
+    ["headache: pain in the head", "chest pain: pressure or tightness in the chest", "rash: skin redness or itching"],
+    mode="multi",
+)
+# {'decision': 'chest pain', 'band': 'auto', 'tags': ['headache', 'chest pain'],
+#  'scores': {'chest pain': 0.76, 'headache': 0.61, 'ask': 0.14, 'rash': 0.0}}
+```
+
+Many texts at once: `m.scores_many(texts, options)`. Command line:
+
+```
+sb-decision "It doesn't work on my machine." -o "crash: the app exits or hangs" -o "slow: works but takes too long" -o "question: asks how to do something"
+```
+
+## Access
+
+The package is public; the weights are in private preview. `SBDecision.load()` raises `SBAccessError` with contact
+details if your Hugging Face token does not have access. To try it on your own option set, contact Sarathi
+Balakrishnan at https://huggingface.co/sarathi-balakrishnan .
+
+## Writing options
+
+- One line per option, saying when to choose it, with the boundary to its neighbours explicit.
+- Describe what the text says, not what you will do about it. "chest pressure with sweating or arm tingling" scores;
+  "tell the caller to call 911" does not.
+- Keep `ask` in the list whenever the action is irreversible.
+- Up to 12 options, 1,024 tokens, English. It decides from the text, not the world: no lookups, no arithmetic.
+
+## Results
+
+Measured on 4,745 messages from five domains never seen in training or validation.
+
+| model | params | held-out top-1 vs teacher | ask recall | false ask | calibration error (ECE) |
+|---|---|---|---|---|---|
+| base | 151M | 0.765 | 0.752 | 0.024 | 0.030 |
+| large | 399M | 0.796 | 0.822 | 0.021 | 0.072 |
+
+The teacher's own self-agreement on the same messages is 0.883.
+
+## License
+
+Apache-2.0. See NOTICE.

sb_decision-0.2.0/README.md

@@ -0,0 +1,82 @@
+# sb-decision
+
+The SB Decision Model, by Sarathi Balakrishnan. Give it a text and your own list of options, each written as
+`name: when to choose it`, and it returns a calibrated probability per option in one pass. Nothing is retrained per
+use case: the option definitions are the policy. A reserved `ask` option is trained to win when the deciding fact is
+missing from the text, so irreversible actions go to a person instead of being guessed.
+
+Not an LLM. It never generates text. It is the router, guardrail, or triage layer in front of one.
+
+## Install
+
+```
+pip install sb-decision
+```
+
+## Use
+
+```python
+from sb_decision import SBDecision
+
+m = SBDecision.load()  # sarathi-balakrishnan/browser-decision-model-base
+m.decide(
+    "Package arrived broken, second time this month. I want a manager.",
+    [
+        "refund: customer wants money back for a damaged, wrong, or missing item",
+        "replace: customer wants the product again",
+        "escalate: angry, threatens to leave, or asks for a manager",
+    ],
+)
+# {'decision': 'escalate', 'band': 'auto', 'scores': {'escalate': 0.90, 'ask': 0.09, 'refund': 0.01, 'replace': 0.01}}
+```
+
+`band` is `auto` when the top score is at or above `auto=0.7`, `review` at or above `review=0.4`, otherwise
+`abstain`. If `ask` wins it is always `abstain`. Both thresholds are arguments; they are statements about the error
+rate you accept, not model constants.
+
+Independent tags, when several options can apply at once:
+
+```python
+m.decide(
+    "Headache since morning and now my chest feels tight.",
+    ["headache: pain in the head", "chest pain: pressure or tightness in the chest", "rash: skin redness or itching"],
+    mode="multi",
+)
+# {'decision': 'chest pain', 'band': 'auto', 'tags': ['headache', 'chest pain'],
+#  'scores': {'chest pain': 0.76, 'headache': 0.61, 'ask': 0.14, 'rash': 0.0}}
+```
+
+Many texts at once: `m.scores_many(texts, options)`. Command line:
+
+```
+sb-decision "It doesn't work on my machine." -o "crash: the app exits or hangs" -o "slow: works but takes too long" -o "question: asks how to do something"
+```
+
+## Access
+
+The package is public; the weights are in private preview. `SBDecision.load()` raises `SBAccessError` with contact
+details if your Hugging Face token does not have access. To try it on your own option set, contact Sarathi
+Balakrishnan at https://huggingface.co/sarathi-balakrishnan .
+
+## Writing options
+
+- One line per option, saying when to choose it, with the boundary to its neighbours explicit.
+- Describe what the text says, not what you will do about it. "chest pressure with sweating or arm tingling" scores;
+  "tell the caller to call 911" does not.
+- Keep `ask` in the list whenever the action is irreversible.
+- Up to 12 options, 1,024 tokens, English. It decides from the text, not the world: no lookups, no arithmetic.
+
+## Results
+
+Measured on 4,745 messages from five domains never seen in training or validation.
+
+| model | params | held-out top-1 vs teacher | ask recall | false ask | calibration error (ECE) |
+|---|---|---|---|---|---|
+| base | 151M | 0.765 | 0.752 | 0.024 | 0.030 |
+| large | 399M | 0.796 | 0.822 | 0.021 | 0.072 |
+
+The teacher's own self-agreement on the same messages is 0.883.
+
+## License
+
+Apache-2.0. See NOTICE.

sb_decision-0.2.0/pyproject.toml

@@ -0,0 +1,34 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "sb-decision"
+version = "0.2.0"
+description = "SB Decision Model: score your own options for any text in one pass, with a trained-in ask. By Sarathi Balakrishnan."
+readme = "README.md"
+authors = [{ name = "Sarathi Balakrishnan" }]
+license = { text = "Apache-2.0" }
+requires-python = ">=3.10"
+keywords = ["decision", "routing", "triage", "classification", "zero-shot", "guardrail"]
+classifiers = [
+  "Development Status :: 4 - Beta",
+  "Intended Audience :: Developers",
+  "License :: OSI Approved :: Apache Software License",
+  "Programming Language :: Python :: 3",
+  "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = ["torch>=2.2", "transformers>=4.48", "huggingface_hub>=0.24"]
+
+[project.urls]
+Model = "https://huggingface.co/sarathi-balakrishnan/browser-decision-model-base"
+Author = "https://huggingface.co/sarathi-balakrishnan"
+
+[project.scripts]
+sb-decision = "sb_decision.cli:main"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools]
+license-files = ["LICENSE", "NOTICE"]

sb_decision-0.2.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

sb_decision-0.2.0/src/sb_decision/__init__.py

@@ -0,0 +1,22 @@
+"""SB Decision Model by Sarathi Balakrishnan: your options, scored in one pass, with a trained-in ask.
+
+from sb_decision import SBDecision
+m = SBDecision.load()
+m.decide("Package arrived broken, I want a manager.",
+         ["refund: customer wants money back", "escalate: angry or asks for a manager"])
+"""
+
+from sb_decision.core import ASK, DEFAULT_MODEL, SBAccessError, SBDecision, parse_options
+from sb_decision.modeling import SBDecisionConfig, SBDecisionModel
+
+__version__ = "0.2.0"
+__all__ = [
+    "ASK",
+    "DEFAULT_MODEL",
+    "SBAccessError",
+    "SBDecision",
+    "SBDecisionConfig",
+    "SBDecisionModel",
+    "parse_options",
+    "__version__",
+]

sb_decision-0.2.0/src/sb_decision/cli.py

@@ -0,0 +1,38 @@
+"""sb-decision "text" -o "refund: customer wants money back" -o "escalate: asks for a manager" """
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+
+from sb_decision.core import DEFAULT_MODEL, SBAccessError, SBDecision
+
+
+def main(argv: list[str] | None = None) -> int:
+    ap = argparse.ArgumentParser(
+        prog="sb-decision", description="Score your own options for a text with the SB Decision Model."
+    )
+    ap.add_argument("text")
+    ap.add_argument("-o", "--option", action="append", required=True, help="'name: when to choose it' (repeat)")
+    ap.add_argument(
+        "--mode", choices=["single", "multi"], default="single", help="single = pick one, multi = independent tags"
+    )
+    ap.add_argument("--no-ask", action="store_true", help="do not add the trained 'ask' option")
+    ap.add_argument("--auto", type=float, default=0.7)
+    ap.add_argument("--review", type=float, default=0.4)
+    ap.add_argument("--model", default=DEFAULT_MODEL, help="Hugging Face repo id or local folder")
+    ap.add_argument("--device", default="cpu")
+    a = ap.parse_args(argv)
+    try:
+        m = SBDecision.load(a.model, device=a.device)
+    except SBAccessError as e:
+        print(str(e), file=sys.stderr)
+        return 2
+    r = m.decide(a.text, a.option, mode=a.mode, include_ask=not a.no_ask, auto=a.auto, review=a.review)
+    print(json.dumps(r, indent=1))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

sb_decision-0.2.0/src/sb_decision/core.py

@@ -0,0 +1,135 @@
+"""SBDecision: load the model once, then decide(text, options)."""
+
+from __future__ import annotations
+
+import torch
+from huggingface_hub.errors import GatedRepoError, HfHubHTTPError, RepositoryNotFoundError
+from transformers import AutoTokenizer
+
+from sb_decision.modeling import LABEL_TOKEN, SEP_TOKEN, SBDecisionModel
+
+DEFAULT_MODEL = "sarathi-balakrishnan/browser-decision-model-base"
+ASK = {"name": "ask", "definition": "not enough information to decide"}
+AUTHOR_PAGE = "https://huggingface.co/sarathi-balakrishnan"
+
+
+class SBAccessError(RuntimeError):
+    """The package installed fine, but the weights are not reachable with this token."""
+
+
+def parse_options(options: list[str | dict]) -> list[dict]:
+    """'name: when to choose it' strings or {'name', 'definition'} dicts -> dicts. Names are lower-cased."""
+    out: list[dict] = []
+    for o in options:
+        if isinstance(o, str):
+            name, _, definition = o.partition(":")
+            out.append({"name": name.strip().lower(), "definition": definition.strip()})
+        else:
+            out.append({"name": str(o["name"]).strip().lower(), "definition": str(o.get("definition", "")).strip()})
+    if any(not o["name"] for o in out):
+        raise ValueError("every option needs a name")
+    return out
+
+
+def band(decision: str, p: float, auto: float, review: float) -> str:
+    if decision == ASK["name"]:
+        return "abstain"
+    return "auto" if p >= auto else "review" if p >= review else "abstain"
+
+
+class SBDecision:
+    """One loaded model. ``decide`` for a verdict with a band, ``scores`` for the raw calibrated probabilities."""
+
+    def __init__(self, model: SBDecisionModel, tokenizer, device: str = "cpu") -> None:
+        self.model = model.to(device).eval()
+        self.tokenizer = tokenizer
+        self.device = device
+
+    @classmethod
+    def load(
+        cls, model: str = DEFAULT_MODEL, device: str = "cpu", token: str | None = None, revision: str | None = None
+    ) -> SBDecision:
+        """``model`` is a Hugging Face repo id or a local folder. The weights are in private preview: loading needs a
+        Hugging Face token with access, otherwise SBAccessError says how to get it."""
+        try:
+            m = SBDecisionModel.from_pretrained(model, token=token, revision=revision)
+            tok = AutoTokenizer.from_pretrained(model, token=token, revision=revision)
+        except (GatedRepoError, RepositoryNotFoundError) as e:
+            raise SBAccessError(_access_message(model)) from e
+        except HfHubHTTPError as e:
+            if e.response is not None and e.response.status_code in (401, 403, 404):
+                raise SBAccessError(_access_message(model)) from e
+            raise
+        except OSError as e:
+            if "401" in str(e) or "403" in str(e) or "404" in str(e) or "not a valid model identifier" in str(e):
+                raise SBAccessError(_access_message(model)) from e
+            raise
+        return cls(m, tok, device)
+
+    @property
+    def max_options(self) -> int:
+        return int(getattr(self.model.config, "max_options", 12))
+
+    def build_input(self, text: str, options: list[dict]) -> str:
+        labels = [f"{o['name']}: {o['definition']}" if o["definition"] else o["name"] for o in options]
+        return "".join(LABEL_TOKEN + lab for lab in labels) + SEP_TOKEN + text
+
+    @torch.no_grad()
+    def scores_many(
+        self,
+        texts: list[str],
+        options: list[str | dict],
+        mode: str = "single",
+        include_ask: bool = True,
+        max_len: int = 1024,
+        batch_size: int = 16,
+    ) -> list[dict[str, float]]:
+        opts = parse_options(options)
+        if include_ask and all(o["name"] != ASK["name"] for o in opts):
+            opts = [*opts, ASK]
+        if not 2 <= len(opts) <= self.max_options:
+            raise ValueError(f"need 2..{self.max_options} options including ask, got {len(opts)}")
+        if mode not in ("single", "multi"):
+            raise ValueError("mode is 'single' (pick one, softmax) or 'multi' (independent, sigmoid)")
+        out: list[dict[str, float]] = []
+        for i in range(0, len(texts), batch_size):
+            chunk = [self.build_input(t, opts) for t in texts[i : i + batch_size]]
+            enc = self.tokenizer(chunk, truncation=True, max_length=max_len, padding="longest", return_tensors="pt")
+            enc = {k: v.to(self.device) for k, v in enc.items()}
+            z = self.model(**enc, max_num_classes=len(opts), calibrated=True).logits.float()
+            p = torch.softmax(z, -1) if mode == "single" else torch.sigmoid(z)
+            out.extend({o["name"]: round(float(row[k]), 4) for k, o in enumerate(opts)} for row in p.cpu())
+        return out
+
+    def scores(self, text: str, options: list[str | dict], **kw) -> dict[str, float]:
+        return self.scores_many([text], options, **kw)[0]
+
+    def decide(
+        self,
+        text: str,
+        options: list[str | dict],
+        mode: str = "single",
+        include_ask: bool = True,
+        auto: float = 0.7,
+        review: float = 0.4,
+        **kw,
+    ) -> dict:
+        """{'decision', 'band' (auto | review | abstain), 'scores'}; in multi mode also 'tags' (every option >= 0.5)."""
+        s = self.scores(text, options, mode=mode, include_ask=include_ask, **kw)
+        top = max(s, key=s.get)
+        r = {
+            "decision": top,
+            "band": band(top, s[top], auto, review),
+            "scores": dict(sorted(s.items(), key=lambda kv: -kv[1])),
+        }
+        if mode == "multi":
+            r["tags"] = [n for n, p in s.items() if p >= 0.5 and n != ASK["name"]]
+        return r
+
+
+def _access_message(model: str) -> str:
+    return (
+        f"sb-decision is installed, but the model weights at '{model}' are not reachable with this Hugging Face token. "
+        "The SB Decision Model is in private preview: to get access, contact Sarathi Balakrishnan at "
+        f"{AUTHOR_PAGE} . Once granted, log in with `hf auth login` or pass token=... to SBDecision.load()."
+    )

sb_decision-0.2.0/src/sb_decision/modeling.py

@@ -0,0 +1,149 @@
+"""SB Decision Model: the inference architecture, self-contained (torch + transformers only).
+
+This file ships inside the Hugging Face repo, so the model loads with
+``AutoModel.from_pretrained(repo, trust_remote_code=True)`` and no other package.
+
+Architecture: a ModernBERT encoder reads ``<<LABEL>>name: definition … <<SEP>>text`` in one pass; the hidden
+state at each ``<<LABEL>>`` token (projected) is dotted with the projected [CLS] state to give one logit per option.
+The state-dict layout matches the training checkpoint (``model.encoder_model.*``, ``model.text_projector.*``,
+``model.classes_projector.*``), so fine-tuned weights load unchanged.
+
+Copyright 2026 Sarathi Balakrishnan. Apache-2.0. See NOTICE for the encoder and recipe this builds on.
+"""
+
+from __future__ import annotations
+
+import torch
+from torch import nn
+from transformers import AutoConfig, AutoModel, PretrainedConfig, PreTrainedModel
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import SequenceClassifierOutput
+
+LABEL_TOKEN = "<<LABEL>>"
+SEP_TOKEN = "<<SEP>>"
+
+
+class SBDecisionConfig(PretrainedConfig):
+    model_type = "sb_decision"
+
+    def __init__(
+        self,
+        encoder_config: dict | PretrainedConfig | None = None,
+        hidden_size: int = 768,
+        class_token_index: int = 50368,
+        text_token_index: int = 50369,
+        calibration_temperature: float = 1.0,
+        projector_hidden_act: str = "gelu",
+        dropout: float = 0.1,
+        max_options: int = 12,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        if isinstance(encoder_config, dict):
+            encoder_config = AutoConfig.for_model(**encoder_config)
+        self.encoder_config = encoder_config
+        self.hidden_size = hidden_size
+        self.class_token_index = class_token_index
+        self.text_token_index = text_token_index
+        self.calibration_temperature = calibration_temperature  # "temperature" alone is reserved for generation
+        self.projector_hidden_act = projector_hidden_act
+        self.dropout = dropout
+        self.max_options = max_options
+
+    def to_dict(self) -> dict:
+        d = super().to_dict()
+        if self.encoder_config is not None:
+            d["encoder_config"] = self.encoder_config.to_dict()
+        return d
+
+
+class Projector(nn.Module):
+    def __init__(self, enc_dim: int, hidden: int, act: str, dropout: float) -> None:
+        super().__init__()
+        self.linear_1 = nn.Linear(enc_dim, hidden)
+        self.act = ACT2FN[act]
+        self.dropout = nn.Dropout(dropout)
+        self.linear_2 = nn.Linear(hidden, enc_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear_2(self.dropout(self.act(self.linear_1(x))))
+
+
+class DotScorer(nn.Module):
+    def forward(self, text: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
+        return torch.einsum("bd,bkd->bk", text, labels)
+
+
+class SBDecisionCore(nn.Module):
+    def __init__(self, cfg: SBDecisionConfig) -> None:
+        super().__init__()
+        enc_dim = cfg.encoder_config.hidden_size
+        self.encoder_model = AutoModel.from_config(cfg.encoder_config)
+        self.text_projector = Projector(enc_dim, cfg.hidden_size, cfg.projector_hidden_act, cfg.dropout)
+        self.classes_projector = Projector(enc_dim, cfg.hidden_size, cfg.projector_hidden_act, cfg.dropout)
+        self.logit_scale = nn.Parameter(torch.tensor(1.0))  # present in the checkpoint, unused at inference
+        self.scorer = DotScorer()
+
+
+def label_positions_from_ids(input_ids: torch.Tensor, label_token_id: int, k: int | None = None) -> torch.Tensor:
+    """[B, K] index of each <<LABEL>> token, -1 padded."""
+    mask = input_ids == label_token_id
+    k = k or int(mask.sum(-1).max().item())
+    pos = torch.full((input_ids.shape[0], k), -1, dtype=torch.long, device=input_ids.device)
+    for b in range(input_ids.shape[0]):
+        idx = mask[b].nonzero().flatten()[:k]
+        pos[b, : idx.numel()] = idx
+    return pos
+
+
+class SBDecisionModel(PreTrainedModel):
+    config_class = SBDecisionConfig
+    base_model_prefix = "model"
+
+    def __init__(self, config: SBDecisionConfig) -> None:
+        super().__init__(config)
+        self.model = SBDecisionCore(config)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        label_positions: torch.Tensor | None = None,
+        max_num_classes: int | None = None,
+        calibrated: bool = False,
+        **kwargs,
+    ) -> SequenceClassifierOutput:
+        """logits[b, k] for option k of row b; padded slots are -1e4. ``calibrated`` divides by the temperature."""
+        m = self.model
+        hidden = m.encoder_model(input_ids=input_ids, attention_mask=attention_mask)[0]  # [B, S, D]
+        if label_positions is None:
+            label_positions = label_positions_from_ids(input_ids, self.config.class_token_index, max_num_classes)
+        valid = (label_positions >= 0).to(hidden.dtype)
+        idx = label_positions.clamp(min=0).unsqueeze(-1).expand(-1, -1, hidden.shape[-1])
+        labels = m.classes_projector(torch.gather(hidden, 1, idx))  # [B, K, D]
+        text = m.text_projector(hidden[:, 0, :])  # [B, D]
+        logits = m.scorer(text, labels) * valid + (valid - 1.0) * 1e4
+        if calibrated:
+            logits = logits / self.config.calibration_temperature
+        return SequenceClassifierOutput(logits=logits)
+
+    @staticmethod
+    def build_input(text: str, options: list[dict]) -> str:
+        labels = [f"{o['name']}: {o['definition']}" if o.get("definition") else o["name"] for o in options]
+        return "".join(LABEL_TOKEN + label for label in labels) + SEP_TOKEN + text
+
+    @torch.no_grad()
+    def score(self, tokenizer, text: str, options: list[dict], mode: str = "single", max_len: int = 1024) -> dict:
+        """Calibrated probability per option. mode "single" = softmax (pick one), "multi" = sigmoid (independent)."""
+        if not 2 <= len(options) <= self.config.max_options:
+            raise ValueError(f"2..{self.config.max_options} options, got {len(options)}")
+        enc = tokenizer(self.build_input(text, options), truncation=True, max_length=max_len, return_tensors="pt")
+        enc = {k: v.to(self.device) for k, v in enc.items()}
+        z = self(**enc, max_num_classes=len(options), calibrated=True).logits[0].float()
+        p = torch.softmax(z, -1) if mode == "single" else torch.sigmoid(z)
+        return {o["name"]: round(float(p[i]), 4) for i, o in enumerate(options)}
+
+
+AutoConfig.register("sb_decision", SBDecisionConfig, exist_ok=True)
+AutoModel.register(SBDecisionConfig, SBDecisionModel, exist_ok=True)

sb_decision-0.2.0/src/sb_decision.egg-info/PKG-INFO

@@ -0,0 +1,105 @@
+Metadata-Version: 2.4
+Name: sb-decision
+Version: 0.2.0
+Summary: SB Decision Model: score your own options for any text in one pass, with a trained-in ask. By Sarathi Balakrishnan.
+Author: Sarathi Balakrishnan
+License: Apache-2.0
+Project-URL: Model, https://huggingface.co/sarathi-balakrishnan/browser-decision-model-base
+Project-URL: Author, https://huggingface.co/sarathi-balakrishnan
+Keywords: decision,routing,triage,classification,zero-shot,guardrail
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: torch>=2.2
+Requires-Dist: transformers>=4.48
+Requires-Dist: huggingface_hub>=0.24
+Dynamic: license-file
+
+# sb-decision
+
+The SB Decision Model, by Sarathi Balakrishnan. Give it a text and your own list of options, each written as
+`name: when to choose it`, and it returns a calibrated probability per option in one pass. Nothing is retrained per
+use case: the option definitions are the policy. A reserved `ask` option is trained to win when the deciding fact is
+missing from the text, so irreversible actions go to a person instead of being guessed.
+
+Not an LLM. It never generates text. It is the router, guardrail, or triage layer in front of one.
+
+## Install
+
+```
+pip install sb-decision
+```
+
+## Use
+
+```python
+from sb_decision import SBDecision
+
+m = SBDecision.load()  # sarathi-balakrishnan/browser-decision-model-base
+m.decide(
+    "Package arrived broken, second time this month. I want a manager.",
+    [
+        "refund: customer wants money back for a damaged, wrong, or missing item",
+        "replace: customer wants the product again",
+        "escalate: angry, threatens to leave, or asks for a manager",
+    ],
+)
+# {'decision': 'escalate', 'band': 'auto', 'scores': {'escalate': 0.90, 'ask': 0.09, 'refund': 0.01, 'replace': 0.01}}
+```
+
+`band` is `auto` when the top score is at or above `auto=0.7`, `review` at or above `review=0.4`, otherwise
+`abstain`. If `ask` wins it is always `abstain`. Both thresholds are arguments; they are statements about the error
+rate you accept, not model constants.
+
+Independent tags, when several options can apply at once:
+
+```python
+m.decide(
+    "Headache since morning and now my chest feels tight.",
+    ["headache: pain in the head", "chest pain: pressure or tightness in the chest", "rash: skin redness or itching"],
+    mode="multi",
+)
+# {'decision': 'chest pain', 'band': 'auto', 'tags': ['headache', 'chest pain'],
+#  'scores': {'chest pain': 0.76, 'headache': 0.61, 'ask': 0.14, 'rash': 0.0}}
+```
+
+Many texts at once: `m.scores_many(texts, options)`. Command line:
+
+```
+sb-decision "It doesn't work on my machine." -o "crash: the app exits or hangs" -o "slow: works but takes too long" -o "question: asks how to do something"
+```
+
+## Access
+
+The package is public; the weights are in private preview. `SBDecision.load()` raises `SBAccessError` with contact
+details if your Hugging Face token does not have access. To try it on your own option set, contact Sarathi
+Balakrishnan at https://huggingface.co/sarathi-balakrishnan .
+
+## Writing options
+
+- One line per option, saying when to choose it, with the boundary to its neighbours explicit.
+- Describe what the text says, not what you will do about it. "chest pressure with sweating or arm tingling" scores;
+  "tell the caller to call 911" does not.
+- Keep `ask` in the list whenever the action is irreversible.
+- Up to 12 options, 1,024 tokens, English. It decides from the text, not the world: no lookups, no arithmetic.
+
+## Results
+
+Measured on 4,745 messages from five domains never seen in training or validation.
+
+| model | params | held-out top-1 vs teacher | ask recall | false ask | calibration error (ECE) |
+|---|---|---|---|---|---|
+| base | 151M | 0.765 | 0.752 | 0.024 | 0.030 |
+| large | 399M | 0.796 | 0.822 | 0.021 | 0.072 |
+
+The teacher's own self-agreement on the same messages is 0.883.
+
+## License
+
+Apache-2.0. See NOTICE.

sb_decision-0.2.0/src/sb_decision.egg-info/SOURCES.txt

@@ -0,0 +1,15 @@
+LICENSE
+NOTICE
+README.md
+pyproject.toml
+src/sb_decision/__init__.py
+src/sb_decision/cli.py
+src/sb_decision/core.py
+src/sb_decision/modeling.py
+src/sb_decision.egg-info/PKG-INFO
+src/sb_decision.egg-info/SOURCES.txt
+src/sb_decision.egg-info/dependency_links.txt
+src/sb_decision.egg-info/entry_points.txt
+src/sb_decision.egg-info/requires.txt
+src/sb_decision.egg-info/top_level.txt
+tests/test_sb_decision.py

sb_decision-0.2.0/src/sb_decision.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

sb_decision-0.2.0/src/sb_decision.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+sb-decision = sb_decision.cli:main

sb_decision-0.2.0/src/sb_decision.egg-info/requires.txt

@@ -0,0 +1,3 @@
+torch>=2.2
+transformers>=4.48
+huggingface_hub>=0.24

sb_decision-0.2.0/src/sb_decision.egg-info/top_level.txt

@@ -0,0 +1 @@
+sb_decision

sb_decision-0.2.0/tests/test_sb_decision.py

@@ -0,0 +1,85 @@
+"""sb-decision on a tiny random model: option parsing, bands, modes, batching, save/load, CLI, access error."""
+
+from __future__ import annotations
+
+import json
+
+import pytest
+import torch
+from sb_decision import ASK, SBAccessError, SBDecision, SBDecisionConfig, SBDecisionModel, parse_options
+from sb_decision.cli import main
+from sb_decision.core import band
+from tokenizers import Tokenizer, models, pre_tokenizers
+from transformers import PreTrainedTokenizerFast
+
+LABEL, SEP = 60, 61
+
+
+@pytest.fixture(scope="module")
+def tiny(tmp_path_factory):
+    torch.manual_seed(0)
+    enc = {
+        "model_type": "modernbert", "hidden_size": 32, "intermediate_size": 64, "num_hidden_layers": 2,
+        "num_attention_heads": 4, "vocab_size": 64, "max_position_embeddings": 128, "local_attention": 8,
+        "global_attn_every_n_layers": 1, "pad_token_id": 0, "bos_token_id": 1, "eos_token_id": 2,
+        "cls_token_id": 1, "sep_token_id": 2,
+    }  # fmt: skip
+    cfg = SBDecisionConfig(
+        encoder_config=enc, hidden_size=16, class_token_index=LABEL, text_token_index=SEP, calibration_temperature=0.5
+    )
+    model = SBDecisionModel(cfg).eval()
+    vocab = {"[PAD]": 0, "[CLS]": 1, "[SEP]": 2, "[UNK]": 3, **{f"w{i}": i for i in range(4, 60)}}
+    vocab.update({"<<LABEL>>": LABEL, "<<SEP>>": SEP})
+    t = Tokenizer(models.WordLevel(vocab, unk_token="[UNK]"))
+    t.pre_tokenizer = pre_tokenizers.Whitespace()
+    tok = PreTrainedTokenizerFast(
+        tokenizer_object=t, pad_token="[PAD]", cls_token="[CLS]", sep_token="[SEP]", unk_token="[UNK]"
+    )
+    tok.add_special_tokens({"additional_special_tokens": ["<<LABEL>>", "<<SEP>>"]})
+    d = tmp_path_factory.mktemp("tiny")
+    model.save_pretrained(d)
+    tok.save_pretrained(d)
+    return str(d)
+
+
+def test_parse_options():
+    assert parse_options(["Refund: money back", {"name": "Escalate", "definition": "manager"}, "ask"]) == [
+        {"name": "refund", "definition": "money back"},
+        {"name": "escalate", "definition": "manager"},
+        {"name": "ask", "definition": ""},
+    ]
+    with pytest.raises(ValueError):
+        parse_options([": no name"])
+
+
+def test_band():
+    assert band("refund", 0.9, 0.7, 0.4) == "auto"
+    assert band("refund", 0.5, 0.7, 0.4) == "review"
+    assert band("refund", 0.1, 0.7, 0.4) == "abstain"
+    assert band("ask", 0.99, 0.7, 0.4) == "abstain"
+
+
+def test_decide_shapes_and_modes(tiny):
+    m = SBDecision.load(tiny)
+    opts = ["w10: w11 w12", "w13: w14"]
+    r = m.decide("w20 w21 w22", opts)
+    assert set(r) == {"decision", "band", "scores"} and set(r["scores"]) == {"w10", "w13", ASK["name"]}
+    assert abs(sum(r["scores"].values()) - 1) < 0.01  # softmax
+    r2 = m.decide("w20 w21 w22", opts, mode="multi", include_ask=False)
+    assert set(r2["scores"]) == {"w10", "w13"} and "tags" in r2
+    many = m.scores_many(["w20", "w21 w22", "w23"], opts, batch_size=2)
+    assert len(many) == 3 and many[0] == m.scores("w20", opts)
+    with pytest.raises(ValueError):
+        m.decide("w20", ["w10: w11"], include_ask=False)
+
+
+def test_cli(tiny, capsys):
+    assert main(["w20 w21", "-o", "w10: w11", "-o", "w13: w14", "--model", tiny]) == 0
+    out = json.loads(capsys.readouterr().out)
+    assert out["band"] in ("auto", "review", "abstain")
+
+
+def test_access_error_message():
+    with pytest.raises(SBAccessError) as e:
+        SBDecision.load("sarathi-balakrishnan/this-repo-does-not-exist-xyz", token="hf_invalid")
+    assert "private preview" in str(e.value) and "sarathi-balakrishnan" in str(e.value)