typeseg 0.2.5__tar.gz → 0.2.7__tar.gz

{typeseg-0.2.5 → typeseg-0.2.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: typeseg
-Version: 0.2.5
+Version: 0.2.7
 Summary: Fine-grained, character-level content-type segmentation for textual inputs (U-Net + Mamba).
 Author: Martin Dallinger
 License-Expression: Apache-2.0
@@ -33,7 +33,9 @@ Dynamic: license-file
 `typeseg` labels every character position of a text with one of 35 content types
 (`html`, `css`, `javascript_typescript`, `python`, `powershell`, `encoding_base64`, …),
 recovering the internal structure of mixed, malformed, or convention-breaking inputs.
-The only runtime dependency is `numpy`.
+Runtime dependencies are just `numpy` and `onnxruntime` — both pulled by `pip install
+typeseg`, so the fast ONNX CPU backend works out of the box at arbitrary input length,
+no GPU required. (A pure-numpy fallback also ships, used when onnxruntime is absent.)
 
 <p align="center">
   <img src="https://typeseg.martin-dallinger.me/images/llm_segmentation.png" alt="Correct model output on a heavily mangled HTML input with a hidden shell payload" width="460">
@@ -153,7 +155,7 @@ honestly.
 
 **Per-model device (auto).** The two models reach the GPU by different routes:
 
-- `fast()` (U-Net) runs on **onnxruntime-gpu** — ~3–4× faster on GPU; ~140k chars/s.
+- `fast()` (U-Net) runs on **onnxruntime-gpu** — ~2× faster than CPU end-to-end; ~140k chars/s.
 - `precise()` (Mamba) runs its selective-scan through a **custom CUDA scan kernel**
   (a CuPy `RawKernel`, `cupy-cuda12x`), reaching ~58k tokens/s for the raw forward —
   edging out the research model's JAX `associative_scan` (54.9k tok/s). The selective-scan
@@ -210,6 +212,7 @@ typeseg.backend_info()
 #  'fast_providers': ['CUDAExecutionProvider', 'CPUExecutionProvider'],
 #  'precise_providers': ['CuPyCUDA:NVIDIA GeForce RTX 5070 Laptop GPU']}
 
+text = "<div>hi</div>\n.btn { color: red; }\nalert('x');"
 typeseg.fast(text)      # U-Net on onnxruntime-gpu  (~140k chars/s)
 typeseg.precise(text)   # Mamba on the CuPy scan    (~58k tokens/s raw forward)
 ```
@@ -281,3 +284,14 @@ result = precise(text, raw)
 
 The trade-off is noisier segments (single-character runs, frayed boundaries); the
 default keeps post-processing on for cleaner, deployment-style output.
+
+## Acknowledgments
+
+This work was carried out with extensive technical feedback from and many informative discussions with
+Dr. Yanick Fratantonio and Dr. Luca Invernizzi from Google Security Research (authors of
+[Magika](https://github.com/google/magika)), who also provided generous access to Google Cloud compute
+resources. Gemini models were used to create and refine the dense segment annotations and to act as the
+active-learning oracle. Thanks also go to Univ.-Prof. Stefan Rass (JKU Secure Systems Group) for his
+guidance, especially regarding parsing and classical computer-science approaches. I am especially
+grateful for his detailed feedback on the accompanying master's thesis and for him supporting this
+academic collaboration.

{typeseg-0.2.5 → typeseg-0.2.7}/README.md

@@ -10,7 +10,9 @@
 `typeseg` labels every character position of a text with one of 35 content types
 (`html`, `css`, `javascript_typescript`, `python`, `powershell`, `encoding_base64`, …),
 recovering the internal structure of mixed, malformed, or convention-breaking inputs.
-The only runtime dependency is `numpy`.
+Runtime dependencies are just `numpy` and `onnxruntime` — both pulled by `pip install
+typeseg`, so the fast ONNX CPU backend works out of the box at arbitrary input length,
+no GPU required. (A pure-numpy fallback also ships, used when onnxruntime is absent.)
 
 <p align="center">
   <img src="https://typeseg.martin-dallinger.me/images/llm_segmentation.png" alt="Correct model output on a heavily mangled HTML input with a hidden shell payload" width="460">
@@ -130,7 +132,7 @@ honestly.
 
 **Per-model device (auto).** The two models reach the GPU by different routes:
 
-- `fast()` (U-Net) runs on **onnxruntime-gpu** — ~3–4× faster on GPU; ~140k chars/s.
+- `fast()` (U-Net) runs on **onnxruntime-gpu** — ~2× faster than CPU end-to-end; ~140k chars/s.
 - `precise()` (Mamba) runs its selective-scan through a **custom CUDA scan kernel**
   (a CuPy `RawKernel`, `cupy-cuda12x`), reaching ~58k tokens/s for the raw forward —
   edging out the research model's JAX `associative_scan` (54.9k tok/s). The selective-scan
@@ -187,6 +189,7 @@ typeseg.backend_info()
 #  'fast_providers': ['CUDAExecutionProvider', 'CPUExecutionProvider'],
 #  'precise_providers': ['CuPyCUDA:NVIDIA GeForce RTX 5070 Laptop GPU']}
 
+text = "<div>hi</div>\n.btn { color: red; }\nalert('x');"
 typeseg.fast(text)      # U-Net on onnxruntime-gpu  (~140k chars/s)
 typeseg.precise(text)   # Mamba on the CuPy scan    (~58k tokens/s raw forward)
 ```
@@ -258,3 +261,14 @@ result = precise(text, raw)
 
 The trade-off is noisier segments (single-character runs, frayed boundaries); the
 default keeps post-processing on for cleaner, deployment-style output.
+
+## Acknowledgments
+
+This work was carried out with extensive technical feedback from and many informative discussions with
+Dr. Yanick Fratantonio and Dr. Luca Invernizzi from Google Security Research (authors of
+[Magika](https://github.com/google/magika)), who also provided generous access to Google Cloud compute
+resources. Gemini models were used to create and refine the dense segment annotations and to act as the
+active-learning oracle. Thanks also go to Univ.-Prof. Stefan Rass (JKU Secure Systems Group) for his
+guidance, especially regarding parsing and classical computer-science approaches. I am especially
+grateful for his detailed feedback on the accompanying master's thesis and for him supporting this
+academic collaboration.

{typeseg-0.2.5 → typeseg-0.2.7}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "typeseg"
-version = "0.2.5"
+version = "0.2.7"
 description = "Fine-grained, character-level content-type segmentation for textual inputs (U-Net + Mamba)."
 readme = "README.md"
 license = "Apache-2.0"

{typeseg-0.2.5 → typeseg-0.2.7}/typeseg.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: typeseg
-Version: 0.2.5
+Version: 0.2.7
 Summary: Fine-grained, character-level content-type segmentation for textual inputs (U-Net + Mamba).
 Author: Martin Dallinger
 License-Expression: Apache-2.0
@@ -33,7 +33,9 @@ Dynamic: license-file
 `typeseg` labels every character position of a text with one of 35 content types
 (`html`, `css`, `javascript_typescript`, `python`, `powershell`, `encoding_base64`, …),
 recovering the internal structure of mixed, malformed, or convention-breaking inputs.
-The only runtime dependency is `numpy`.
+Runtime dependencies are just `numpy` and `onnxruntime` — both pulled by `pip install
+typeseg`, so the fast ONNX CPU backend works out of the box at arbitrary input length,
+no GPU required. (A pure-numpy fallback also ships, used when onnxruntime is absent.)
 
 <p align="center">
   <img src="https://typeseg.martin-dallinger.me/images/llm_segmentation.png" alt="Correct model output on a heavily mangled HTML input with a hidden shell payload" width="460">
@@ -153,7 +155,7 @@ honestly.
 
 **Per-model device (auto).** The two models reach the GPU by different routes:
 
-- `fast()` (U-Net) runs on **onnxruntime-gpu** — ~3–4× faster on GPU; ~140k chars/s.
+- `fast()` (U-Net) runs on **onnxruntime-gpu** — ~2× faster than CPU end-to-end; ~140k chars/s.
 - `precise()` (Mamba) runs its selective-scan through a **custom CUDA scan kernel**
   (a CuPy `RawKernel`, `cupy-cuda12x`), reaching ~58k tokens/s for the raw forward —
   edging out the research model's JAX `associative_scan` (54.9k tok/s). The selective-scan
@@ -210,6 +212,7 @@ typeseg.backend_info()
 #  'fast_providers': ['CUDAExecutionProvider', 'CPUExecutionProvider'],
 #  'precise_providers': ['CuPyCUDA:NVIDIA GeForce RTX 5070 Laptop GPU']}
 
+text = "<div>hi</div>\n.btn { color: red; }\nalert('x');"
 typeseg.fast(text)      # U-Net on onnxruntime-gpu  (~140k chars/s)
 typeseg.precise(text)   # Mamba on the CuPy scan    (~58k tokens/s raw forward)
 ```
@@ -281,3 +284,14 @@ result = precise(text, raw)
 
 The trade-off is noisier segments (single-character runs, frayed boundaries); the
 default keeps post-processing on for cleaner, deployment-style output.
+
+## Acknowledgments
+
+This work was carried out with extensive technical feedback from and many informative discussions with
+Dr. Yanick Fratantonio and Dr. Luca Invernizzi from Google Security Research (authors of
+[Magika](https://github.com/google/magika)), who also provided generous access to Google Cloud compute
+resources. Gemini models were used to create and refine the dense segment annotations and to act as the
+active-learning oracle. Thanks also go to Univ.-Prof. Stefan Rass (JKU Secure Systems Group) for his
+guidance, especially regarding parsing and classical computer-science approaches. I am especially
+grateful for his detailed feedback on the accompanying master's thesis and for him supporting this
+academic collaboration.