tinyjev 0.1.1__tar.gz → 0.1.2__tar.gz

{tinyjev-0.1.1/tinyjev.egg-info → tinyjev-0.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tinyjev
-Version: 0.1.1
+Version: 0.1.2
 Summary: Typed decisions on your laptop: one forward pass, calibrated probabilities, no text generation. MLX or PyTorch.
 Author: Ankit Aglawe
 License: MIT
@@ -36,6 +36,8 @@ Requires-Dist: torch>=2.2; extra == "convert"
 Requires-Dist: safetensors>=0.4; extra == "convert"
 Provides-Extra: demo
 Requires-Dist: pillow>=10; extra == "demo"
+Provides-Extra: modelscope
+Requires-Dist: modelscope>=1.9; extra == "modelscope"
 Provides-Extra: dev
 Requires-Dist: pytest>=8; extra == "dev"
 Requires-Dist: pillow>=10; extra == "dev"
@@ -63,6 +65,13 @@ Dynamic: license-file
   <a href="https://github.com/ankit-aglawe/tinyjev/tree/main/examples">Examples</a>
 </p>
 
+<p>
+  <b>English</b> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.zh-CN.md">简体中文</a> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.ja.md">日本語</a> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.ko.md">한국어</a>
+</p>
+
 </div>
 
 TinyJev answers typed questions about text and returns probabilities, in one forward pass, on your own machine.
@@ -90,12 +99,6 @@ TinyJev is MIT licensed.
 Eight real support tickets, one after another, on a base M1. Three questions per ticket in a
 single forward pass, about 110 ms each. Every number in that recording came from a live run.
 
-Half of them route themselves. The other half do not, and that is the interesting part. Ticket
-six says the order was late, the wrong size, and double charged, so the probability spreads
-across all three teams and confidence falls to 0.14. Below the gate it goes to a person instead
-of being guessed at. A message that turns out to be a compliment does the same thing. A model
-that only ever returned a label would have picked a team for both.
-
 ```bash
 pip install 'tinyjev[mlx,demo]'
 python demos/triage_desk.py --gif demo.gif
@@ -225,7 +228,13 @@ A question is `{"type": "choice" | "noul" | "score", "instructions": ..., "crite
 
 One published model, `tinyjev-0.6b`: a Qwen3-0.6B-Base backbone with a pointer head, trained on public typed-decision data. It is a first release and it is small — on question types unlike anything it has seen, use the confidence gate rather than the raw label. Specialising it on a few hundred of your own labelled examples is the intended path and beats prompt tuning a larger model.
 
-Not yet: multilingual, vision, or label sets in the hundreds.
+It handles other languages better than expected, which we did not train for. Translating 25 cases of
+a held-out English benchmark into Chinese and re-running them scored 22/25 against 20/25 for the
+English originals, with the question schema left in English. That is parity on a small sample, not
+a multilingual claim, and confidence runs flatter in Chinese so fewer cases clear a fixed gate. The
+backbone is Qwen3, which is where this comes from.
+
+Not yet: vision, or label sets in the hundreds.
 
 ## Support this project
 

{tinyjev-0.1.1 → tinyjev-0.1.2}/README.md

@@ -18,6 +18,13 @@
   <a href="https://github.com/ankit-aglawe/tinyjev/tree/main/examples">Examples</a>
 </p>
 
+<p>
+  <b>English</b> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.zh-CN.md">简体中文</a> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.ja.md">日本語</a> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.ko.md">한국어</a>
+</p>
+
 </div>
 
 TinyJev answers typed questions about text and returns probabilities, in one forward pass, on your own machine.
@@ -45,12 +52,6 @@ TinyJev is MIT licensed.
 Eight real support tickets, one after another, on a base M1. Three questions per ticket in a
 single forward pass, about 110 ms each. Every number in that recording came from a live run.
 
-Half of them route themselves. The other half do not, and that is the interesting part. Ticket
-six says the order was late, the wrong size, and double charged, so the probability spreads
-across all three teams and confidence falls to 0.14. Below the gate it goes to a person instead
-of being guessed at. A message that turns out to be a compliment does the same thing. A model
-that only ever returned a label would have picked a team for both.
-
 ```bash
 pip install 'tinyjev[mlx,demo]'
 python demos/triage_desk.py --gif demo.gif
@@ -180,7 +181,13 @@ A question is `{"type": "choice" | "noul" | "score", "instructions": ..., "crite
 
 One published model, `tinyjev-0.6b`: a Qwen3-0.6B-Base backbone with a pointer head, trained on public typed-decision data. It is a first release and it is small — on question types unlike anything it has seen, use the confidence gate rather than the raw label. Specialising it on a few hundred of your own labelled examples is the intended path and beats prompt tuning a larger model.
 
-Not yet: multilingual, vision, or label sets in the hundreds.
+It handles other languages better than expected, which we did not train for. Translating 25 cases of
+a held-out English benchmark into Chinese and re-running them scored 22/25 against 20/25 for the
+English originals, with the question schema left in English. That is parity on a small sample, not
+a multilingual claim, and confidence runs flatter in Chinese so fewer cases clear a fixed gate. The
+backbone is Qwen3, which is where this comes from.
+
+Not yet: vision, or label sets in the hundreds.
 
 ## Support this project
 

{tinyjev-0.1.1 → tinyjev-0.1.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tinyjev"
-version = "0.1.1"
+version = "0.1.2"
 description = "Typed decisions on your laptop: one forward pass, calibrated probabilities, no text generation. MLX or PyTorch."
 readme = "README.md"
 requires-python = ">=3.9"
@@ -34,6 +34,7 @@ mlx = ["mlx>=0.22", "mlx-lm>=0.20"]
 torch = ["torch>=2.2", "transformers>=4.45"]
 convert = ["torch>=2.2", "safetensors>=0.4"]
 demo = ["pillow>=10"]
+modelscope = ["modelscope>=1.9"]
 dev = ["pytest>=8", "pillow>=10", "torch>=2.2", "transformers>=4.45"]
 
 [project.scripts]

{tinyjev-0.1.1 → tinyjev-0.1.2}/tinyjev/agent.py

@@ -14,8 +14,16 @@ TYPE_ALIASES = {"noul": "boolean", "boolean": "boolean", "choice": "choice", "sc
 
 
 def _resolve(model_path, subfolder: Optional[str] = None) -> Path:
+    """A local directory, a Hub repo id, or `modelscope:<org>/<repo>` for mirrors of
+    this model on ModelScope, which is reachable from places the Hub is not."""
     from .registry import resolve
-    repo, sub = resolve(str(model_path))
+    name = str(model_path)
+    if name.startswith("modelscope:"):
+        from modelscope import snapshot_download as ms_download
+        root = Path(ms_download(name.split(":", 1)[1]))
+        return root / subfolder if subfolder else root
+
+    repo, sub = resolve(name)
     sub = subfolder or sub
     p = Path(repo).expanduser()
     if p.exists():

{tinyjev-0.1.1 → tinyjev-0.1.2/tinyjev.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tinyjev
-Version: 0.1.1
+Version: 0.1.2
 Summary: Typed decisions on your laptop: one forward pass, calibrated probabilities, no text generation. MLX or PyTorch.
 Author: Ankit Aglawe
 License: MIT
@@ -36,6 +36,8 @@ Requires-Dist: torch>=2.2; extra == "convert"
 Requires-Dist: safetensors>=0.4; extra == "convert"
 Provides-Extra: demo
 Requires-Dist: pillow>=10; extra == "demo"
+Provides-Extra: modelscope
+Requires-Dist: modelscope>=1.9; extra == "modelscope"
 Provides-Extra: dev
 Requires-Dist: pytest>=8; extra == "dev"
 Requires-Dist: pillow>=10; extra == "dev"
@@ -63,6 +65,13 @@ Dynamic: license-file
   <a href="https://github.com/ankit-aglawe/tinyjev/tree/main/examples">Examples</a>
 </p>
 
+<p>
+  <b>English</b> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.zh-CN.md">简体中文</a> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.ja.md">日本語</a> ·
+  <a href="https://github.com/ankit-aglawe/tinyjev/blob/main/README.ko.md">한국어</a>
+</p>
+
 </div>
 
 TinyJev answers typed questions about text and returns probabilities, in one forward pass, on your own machine.
@@ -90,12 +99,6 @@ TinyJev is MIT licensed.
 Eight real support tickets, one after another, on a base M1. Three questions per ticket in a
 single forward pass, about 110 ms each. Every number in that recording came from a live run.
 
-Half of them route themselves. The other half do not, and that is the interesting part. Ticket
-six says the order was late, the wrong size, and double charged, so the probability spreads
-across all three teams and confidence falls to 0.14. Below the gate it goes to a person instead
-of being guessed at. A message that turns out to be a compliment does the same thing. A model
-that only ever returned a label would have picked a team for both.
-
 ```bash
 pip install 'tinyjev[mlx,demo]'
 python demos/triage_desk.py --gif demo.gif
@@ -225,7 +228,13 @@ A question is `{"type": "choice" | "noul" | "score", "instructions": ..., "crite
 
 One published model, `tinyjev-0.6b`: a Qwen3-0.6B-Base backbone with a pointer head, trained on public typed-decision data. It is a first release and it is small — on question types unlike anything it has seen, use the confidence gate rather than the raw label. Specialising it on a few hundred of your own labelled examples is the intended path and beats prompt tuning a larger model.
 
-Not yet: multilingual, vision, or label sets in the hundreds.
+It handles other languages better than expected, which we did not train for. Translating 25 cases of
+a held-out English benchmark into Chinese and re-running them scored 22/25 against 20/25 for the
+English originals, with the question schema left in English. That is parity on a small sample, not
+a multilingual claim, and confidence runs flatter in Chinese so fewer cases clear a fixed gate. The
+backbone is Qwen3, which is where this comes from.
+
+Not yet: vision, or label sets in the hundreds.
 
 ## Support this project
 

{tinyjev-0.1.1 → tinyjev-0.1.2}/tinyjev.egg-info/requires.txt

@@ -20,6 +20,9 @@ transformers>=4.45
 mlx>=0.22
 mlx-lm>=0.20
 
+[modelscope]
+modelscope>=1.9
+
 [torch]
 torch>=2.2
 transformers>=4.45