hallustream 0.2.0__tar.gz

hallustream-0.2.0/LICENSE

@@ -0,0 +1,201 @@
+                                 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.

hallustream-0.2.0/PKG-INFO

@@ -0,0 +1,131 @@
+Metadata-Version: 2.4
+Name: hallustream
+Version: 0.2.0
+Summary: Flag the facts an open LLM is probably making up, from its hidden states, as it writes.
+Keywords: llm,hallucination,probe,interpretability,qwen3,streaming
+Author: Pranav Sethia
+Author-email: Pranav Sethia <pranav.sethia5@gmail.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Dist: torch>=2.4
+Requires-Dist: transformers>=4.56
+Requires-Dist: safetensors>=0.4
+Requires-Dist: numpy>=1.26
+Requires-Dist: pysbd>=0.3.4
+Requires-Dist: gliner>=0.2.24 ; extra == 'entities'
+Requires-Dist: protobuf>=4 ; extra == 'entities'
+Requires-Python: >=3.10
+Project-URL: Homepage, https://pranav-sethia.github.io/hallustream/
+Project-URL: Repository, https://github.com/pranav-sethia/hallustream
+Provides-Extra: entities
+Description-Content-Type: text/markdown
+
+# HalluStream
+
+[![CI](https://github.com/pranav-sethia/hallustream/actions/workflows/ci.yml/badge.svg)](https://github.com/pranav-sethia/hallustream/actions/workflows/ci.yml)
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/pranav-sethia/hallustream/blob/main/notebooks/quickstart.ipynb)
+
+HalluStream catches a language model making things up while it writes. A small linear probe reads the model's hidden states and flags the names, dates, places and titles it is probably inventing. It doesn't generate extra text or search the web, and it adds about 3–7% to generation time.
+
+The [project page](https://pranav-sethia.github.io/hallustream/) replays real runs with the flags appearing live.
+
+![Qwen3-4B writing a biography; likely made-up facts are highlighted as each sentence is checked](https://raw.githubusercontent.com/pranav-sethia/hallustream/main/docs/img/stream.gif)
+
+*A real, unedited run with Qwen3-4B (recorded with the previous version of the detector). The flags land on invented facts: her birth date, her mother, a brother, and films and years she was never in.*
+
+## Quick start
+
+```bash
+pip install "hallustream[entities] @ git+https://github.com/pranav-sethia/hallustream"
+hallustream "Tell me about A Caribbean Mystery."
+```
+
+That answers the question and highlights the details that are likely made up. There is also `hallustream chat`, `score` for checking an answer the model already wrote, and `batch` for whole files. From Python:
+
+```python
+from hallustream import Detector
+
+det = Detector.from_pretrained("qwen3-4b")
+report = det.generate("Tell me a bio of Kiara Advani.")
+print(report.summary())
+```
+
+It needs a GPU or an Apple Silicon Mac for Qwen3-4B, and Python 3.10 or newer. The [Colab notebook](https://colab.research.google.com/github/pranav-sethia/hallustream/blob/main/notebooks/quickstart.ipynb) runs on the free tier. More in [LIBRARY.md](https://github.com/pranav-sethia/hallustream/blob/main/docs/LIBRARY.md).
+
+## Results
+
+Every test below was written down in [EVAL_PROTOCOL.md](https://github.com/pranav-sethia/hallustream/blob/main/EVAL_PROTOCOL.md) before it was run, choices were made on separate validation data, and each test set was scored once. The details are in [RESULTS.md](https://github.com/pranav-sethia/hallustream/blob/main/docs/RESULTS.md).
+
+### Against the paper it builds on
+
+The probe comes from [Obeso et al. (2025)](https://arxiv.org/abs/2509.03531), who released probes for several models. I ran their released probes through my own pipeline, then changed two things: the layer is picked on validation instead of fixed near the end, and the training is tuned on validation. The model is never changed. On the paper's own data and test sets, scored on exactly the same details as the paper's released probes:
+
+| AUROC on LongFact / LongFact++ | Paper's linear probe | Paper's LoRA probe (fine-tunes the model) | HalluStream |
+|---|---|---|---|
+| Qwen2.5-7B | 0.839 / 0.847 | 0.884 / 0.889 | **0.894 / 0.896** |
+| Llama-3.1-8B | 0.848 / 0.862 | **0.889 / 0.900** | 0.872 / 0.891 |
+
+HalluStream beats the paper's linear probe on both models (+0.050 to +0.055 on Qwen2.5-7B, +0.024 to +0.029 on Llama-3.1-8B), and on Qwen2.5-7B it also beats the probe that fine-tunes the model (+0.010 / +0.008). On Llama-3.1-8B, the model with the most training data in the paper, fine-tuning still wins (by 0.017 / 0.009). On the paper's HealthBench set of medical answers, never used here before, the same holds: 0.874 against 0.849 (linear) and 0.892 (LoRA). Every difference holds at 95% confidence.
+
+### On the model's own writing
+
+The paper trains on other models' answers. HalluStream also adapts the probe to the model's own writing. Tested on Qwen3-4B's texts about people and other subjects drawn fresh from Wikipedia after the detector was frozen:
+
+| AUROC | 437 new people | 498 new subjects |
+|---|---|---|
+| The model's own confidence (max entropy) | 0.660 | 0.684 |
+| The paper's recipe on Qwen3-4B | 0.814 | 0.780 |
+| **HalluStream** | **0.874** | **0.854** |
+
+Comparing only details about the same subject, which removes the easy part of telling obscure subjects from famous ones, HalluStream still leads the model's confidence 0.87 to 0.75 and 0.85 to 0.75. Its risk reads as a probability: of the details it gives 90–100% risk, 96% are unsupported.
+
+### Inside SHARS
+
+[SHARS](https://arxiv.org/abs/2606.03628) (ICML 2026) reduces hallucination by writing an answer one sentence at a time and redoing sentences whose facts look doubtful. Its own detector samples extra answers for every fact. I swapped in HalluStream, changed nothing else, and ran both detectors at several settings on all 182 of SHARS's test people:
+
+![False against true claims per person for each detector and setting](https://raw.githubusercontent.com/pranav-sethia/hallustream/main/docs/img/shars_tradeoff.png)
+
+The two detectors trace about the same trade-off between true and false claims, and HalluStream gets there 4.5–6x faster end to end (34–49 s per biography against 211–222 s), with about 140x less time spent on detection. At the same answer rate it has the same numbers of true and false claims; set more cautiously it goes further, down to 1.7 false claims per person. A first comparison reported 44% fewer false claims than SHARS's detector; that compared a cautious HalluStream setting with a permissive one of SHARS's, and the runs at matched settings above replace it.
+
+## How it was tested
+
+- **Pre-registered:** each test's settings and pass bars were committed before it ran. Experiments that were stopped are kept in the protocol's appendix with their outcomes.
+- **Reproduced first:** the paper's released Qwen2.5-7B probe, run through this pipeline, gives 0.8384 / 0.8469 against the published 0.8383 / 0.8467.
+- **Fresh test data:** the own-writing test used 1,100 people and subjects sampled after the detector was frozen.
+- **A validated fact checker:** labels come from Gemini reading the subject's Wikipedia page. Against FActScore's human labels it gets 89% balanced accuracy on 11,049 facts. A stronger model re-checking its labels blind agreed on 84–91%; it marks made-up details reliably and leans toward accepting doubtful ones.
+
+## Limits
+
+The probe needs the model's hidden states, so it works on open-weight models, and each model needs its own probe; the bundled one is for Qwen3-4B-Instruct-2507. It holds back what the model would make up but can't add facts the model doesn't know; when a source is available, retrieval does that better (an early pilot here reached 96% precision by retrieving the person's Wikipedia page). The own-writing labels come from a fact checker that only reads Wikipedia, and the adapted probe was fitted on that checker's labels; a stronger judge did not confirm its precision gains inside SHARS.
+
+## Repository
+
+- `src/hallustream/`: the library (detector, probe, entity finder, CLI, SHARS plug-in, bundled weights)
+- `src/hallustream/research/`: data, training, evaluation, the fact checker and the analysis scripts
+- `modal_jobs/`: the GPU and batch jobs, run on Modal
+- `configs/`: frozen settings and fixed lists of people and subjects
+- `results/public/`: labels, scores and texts behind every reported number
+- `notebooks/`: the Colab quickstart
+- `tests/`: unit tests and a package install test, run in CI
+
+Data sources and licenses are in [DATA.md](https://github.com/pranav-sethia/hallustream/blob/main/docs/DATA.md).
+
+## Credits
+
+- Obeso et al., [Real-Time Detection of Hallucinated Entities in Long-Form Generation](https://arxiv.org/abs/2509.03531) (2025): the probe method, training data and released probes
+- Li et al., [Building Reliable Long-Form Generation via Hallucination Rejection Sampling](https://arxiv.org/abs/2606.03628) (ICML 2026): SHARS
+- Min et al., [FActScore](https://arxiv.org/abs/2305.14251) (2023): the people and human labels
+
+To cite this project, see [CITATION.cff](https://github.com/pranav-sethia/hallustream/blob/main/CITATION.cff).
+
+## License
+
+Apache-2.0

hallustream-0.2.0/README.md

@@ -0,0 +1,101 @@
+# HalluStream
+
+[![CI](https://github.com/pranav-sethia/hallustream/actions/workflows/ci.yml/badge.svg)](https://github.com/pranav-sethia/hallustream/actions/workflows/ci.yml)
+[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/pranav-sethia/hallustream/blob/main/notebooks/quickstart.ipynb)
+
+HalluStream catches a language model making things up while it writes. A small linear probe reads the model's hidden states and flags the names, dates, places and titles it is probably inventing. It doesn't generate extra text or search the web, and it adds about 3–7% to generation time.
+
+The [project page](https://pranav-sethia.github.io/hallustream/) replays real runs with the flags appearing live.
+
+![Qwen3-4B writing a biography; likely made-up facts are highlighted as each sentence is checked](https://raw.githubusercontent.com/pranav-sethia/hallustream/main/docs/img/stream.gif)
+
+*A real, unedited run with Qwen3-4B (recorded with the previous version of the detector). The flags land on invented facts: her birth date, her mother, a brother, and films and years she was never in.*
+
+## Quick start
+
+```bash
+pip install "hallustream[entities] @ git+https://github.com/pranav-sethia/hallustream"
+hallustream "Tell me about A Caribbean Mystery."
+```
+
+That answers the question and highlights the details that are likely made up. There is also `hallustream chat`, `score` for checking an answer the model already wrote, and `batch` for whole files. From Python:
+
+```python
+from hallustream import Detector
+
+det = Detector.from_pretrained("qwen3-4b")
+report = det.generate("Tell me a bio of Kiara Advani.")
+print(report.summary())
+```
+
+It needs a GPU or an Apple Silicon Mac for Qwen3-4B, and Python 3.10 or newer. The [Colab notebook](https://colab.research.google.com/github/pranav-sethia/hallustream/blob/main/notebooks/quickstart.ipynb) runs on the free tier. More in [LIBRARY.md](https://github.com/pranav-sethia/hallustream/blob/main/docs/LIBRARY.md).
+
+## Results
+
+Every test below was written down in [EVAL_PROTOCOL.md](https://github.com/pranav-sethia/hallustream/blob/main/EVAL_PROTOCOL.md) before it was run, choices were made on separate validation data, and each test set was scored once. The details are in [RESULTS.md](https://github.com/pranav-sethia/hallustream/blob/main/docs/RESULTS.md).
+
+### Against the paper it builds on
+
+The probe comes from [Obeso et al. (2025)](https://arxiv.org/abs/2509.03531), who released probes for several models. I ran their released probes through my own pipeline, then changed two things: the layer is picked on validation instead of fixed near the end, and the training is tuned on validation. The model is never changed. On the paper's own data and test sets, scored on exactly the same details as the paper's released probes:
+
+| AUROC on LongFact / LongFact++ | Paper's linear probe | Paper's LoRA probe (fine-tunes the model) | HalluStream |
+|---|---|---|---|
+| Qwen2.5-7B | 0.839 / 0.847 | 0.884 / 0.889 | **0.894 / 0.896** |
+| Llama-3.1-8B | 0.848 / 0.862 | **0.889 / 0.900** | 0.872 / 0.891 |
+
+HalluStream beats the paper's linear probe on both models (+0.050 to +0.055 on Qwen2.5-7B, +0.024 to +0.029 on Llama-3.1-8B), and on Qwen2.5-7B it also beats the probe that fine-tunes the model (+0.010 / +0.008). On Llama-3.1-8B, the model with the most training data in the paper, fine-tuning still wins (by 0.017 / 0.009). On the paper's HealthBench set of medical answers, never used here before, the same holds: 0.874 against 0.849 (linear) and 0.892 (LoRA). Every difference holds at 95% confidence.
+
+### On the model's own writing
+
+The paper trains on other models' answers. HalluStream also adapts the probe to the model's own writing. Tested on Qwen3-4B's texts about people and other subjects drawn fresh from Wikipedia after the detector was frozen:
+
+| AUROC | 437 new people | 498 new subjects |
+|---|---|---|
+| The model's own confidence (max entropy) | 0.660 | 0.684 |
+| The paper's recipe on Qwen3-4B | 0.814 | 0.780 |
+| **HalluStream** | **0.874** | **0.854** |
+
+Comparing only details about the same subject, which removes the easy part of telling obscure subjects from famous ones, HalluStream still leads the model's confidence 0.87 to 0.75 and 0.85 to 0.75. Its risk reads as a probability: of the details it gives 90–100% risk, 96% are unsupported.
+
+### Inside SHARS
+
+[SHARS](https://arxiv.org/abs/2606.03628) (ICML 2026) reduces hallucination by writing an answer one sentence at a time and redoing sentences whose facts look doubtful. Its own detector samples extra answers for every fact. I swapped in HalluStream, changed nothing else, and ran both detectors at several settings on all 182 of SHARS's test people:
+
+![False against true claims per person for each detector and setting](https://raw.githubusercontent.com/pranav-sethia/hallustream/main/docs/img/shars_tradeoff.png)
+
+The two detectors trace about the same trade-off between true and false claims, and HalluStream gets there 4.5–6x faster end to end (34–49 s per biography against 211–222 s), with about 140x less time spent on detection. At the same answer rate it has the same numbers of true and false claims; set more cautiously it goes further, down to 1.7 false claims per person. A first comparison reported 44% fewer false claims than SHARS's detector; that compared a cautious HalluStream setting with a permissive one of SHARS's, and the runs at matched settings above replace it.
+
+## How it was tested
+
+- **Pre-registered:** each test's settings and pass bars were committed before it ran. Experiments that were stopped are kept in the protocol's appendix with their outcomes.
+- **Reproduced first:** the paper's released Qwen2.5-7B probe, run through this pipeline, gives 0.8384 / 0.8469 against the published 0.8383 / 0.8467.
+- **Fresh test data:** the own-writing test used 1,100 people and subjects sampled after the detector was frozen.
+- **A validated fact checker:** labels come from Gemini reading the subject's Wikipedia page. Against FActScore's human labels it gets 89% balanced accuracy on 11,049 facts. A stronger model re-checking its labels blind agreed on 84–91%; it marks made-up details reliably and leans toward accepting doubtful ones.
+
+## Limits
+
+The probe needs the model's hidden states, so it works on open-weight models, and each model needs its own probe; the bundled one is for Qwen3-4B-Instruct-2507. It holds back what the model would make up but can't add facts the model doesn't know; when a source is available, retrieval does that better (an early pilot here reached 96% precision by retrieving the person's Wikipedia page). The own-writing labels come from a fact checker that only reads Wikipedia, and the adapted probe was fitted on that checker's labels; a stronger judge did not confirm its precision gains inside SHARS.
+
+## Repository
+
+- `src/hallustream/`: the library (detector, probe, entity finder, CLI, SHARS plug-in, bundled weights)
+- `src/hallustream/research/`: data, training, evaluation, the fact checker and the analysis scripts
+- `modal_jobs/`: the GPU and batch jobs, run on Modal
+- `configs/`: frozen settings and fixed lists of people and subjects
+- `results/public/`: labels, scores and texts behind every reported number
+- `notebooks/`: the Colab quickstart
+- `tests/`: unit tests and a package install test, run in CI
+
+Data sources and licenses are in [DATA.md](https://github.com/pranav-sethia/hallustream/blob/main/docs/DATA.md).
+
+## Credits
+
+- Obeso et al., [Real-Time Detection of Hallucinated Entities in Long-Form Generation](https://arxiv.org/abs/2509.03531) (2025): the probe method, training data and released probes
+- Li et al., [Building Reliable Long-Form Generation via Hallucination Rejection Sampling](https://arxiv.org/abs/2606.03628) (ICML 2026): SHARS
+- Min et al., [FActScore](https://arxiv.org/abs/2305.14251) (2023): the people and human labels
+
+To cite this project, see [CITATION.cff](https://github.com/pranav-sethia/hallustream/blob/main/CITATION.cff).
+
+## License
+
+Apache-2.0

hallustream-0.2.0/pyproject.toml

@@ -0,0 +1,91 @@
+[project]
+name = "hallustream"
+version = "0.2.0"
+description = "Flag the facts an open LLM is probably making up, from its hidden states, as it writes."
+readme = "README.md"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+authors = [{ name = "Pranav Sethia", email = "pranav.sethia5@gmail.com" }]
+requires-python = ">=3.10"
+keywords = ["llm", "hallucination", "probe", "interpretability", "qwen3", "streaming"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "torch>=2.4",
+    "transformers>=4.56",
+    "safetensors>=0.4",
+    "numpy>=1.26",
+    "pysbd>=0.3.4",
+]
+
+[project.scripts]
+hallustream = "hallustream.cli:main"
+
+[project.optional-dependencies]
+# protobuf: GLiNER's DeBERTa tokenizer can't be converted without it under transformers 5
+entities = ["gliner>=0.2.24", "protobuf>=4"]
+
+[dependency-groups]
+# Training, evaluation and fact checking (src/hallustream/research, modal_jobs); not published.
+research = [
+    "torch>=2.8",
+    "transformers>=5.20,<6",
+    "huggingface_hub>=1.0",
+    "datasets>=3.0",
+    "scikit-learn>=1.5",
+    "pandas>=2.2",
+    "pyarrow>=17",
+    "wandb>=0.19",
+    "anthropic>=0.60",
+    "matplotlib>=3.10.9",
+    "google-genai>=2.32.0",
+    "gliner>=0.2.24; python_version >= '3.11'",  # its onnxruntime pin has no 3.10 Linux wheel
+    "pysbd>=0.3.4",
+    "sentencepiece>=0.2.2",
+    "protobuf>=7.36.2",
+]
+dev = [
+    "pytest>=8",
+    "hypothesis>=6.100",
+    "ruff>=0.8",
+    "pyright>=1.1.390",
+    "pre-commit>=4",
+    "modal>=1.6",
+]
+
+[project.urls]
+Homepage = "https://pranav-sethia.github.io/hallustream/"
+Repository = "https://github.com/pranav-sethia/hallustream"
+
+[tool.uv.build-backend]
+# The research code (training, evaluation, fact checking) stays in the repo, not in the wheel.
+wheel-exclude = ["hallustream/research/**"]
+
+[build-system]
+requires = ["uv_build>=0.10.7,<0.11.0"]
+build-backend = "uv_build"
+
+[tool.ruff]
+line-length = 100
+target-version = "py310"
+extend-exclude = ["third_party"]
+
+[tool.ruff.lint]
+select = ["E", "F", "I", "B", "UP"]
+
+[tool.pyright]
+include = ["src/hallustream"]
+exclude = ["src/hallustream/research"]
+typeCheckingMode = "basic"
+pythonVersion = "3.10"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

hallustream-0.2.0/src/hallustream/__init__.py

@@ -0,0 +1,18 @@
+"""HalluStream: flag the facts an open LLM is probably making up, from its hidden states."""
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from hallustream.detector import Detector, Entity, Event, Report
+
+__version__ = "0.2.0"
+__all__ = ["Detector", "Entity", "Event", "Report", "__version__"]
+
+
+def __getattr__(name: str):
+    # Loaded on first use, so the research code can import hallustream.* without the library
+    if name in ("Detector", "Entity", "Event", "Report"):
+        from hallustream import detector
+
+        return getattr(detector, name)
+    raise AttributeError(name)