datachain 0.20.1__tar.gz → 0.20.2__tar.gz

{datachain-0.20.1 → datachain-0.20.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: datachain
-Version: 0.20.1
+Version: 0.20.2
 Summary: Wrangle unstructured AI data at scale
 Author-email: Dmitry Petrov <support@dvc.org>
 License-Expression: Apache-2.0

{datachain-0.20.1 → datachain-0.20.2}/examples/llm_and_nlp/hf-dataset-llm-eval.py

@@ -1,3 +1,5 @@
+import os
+
 from huggingface_hub import InferenceClient
 from requests import HTTPError
 
@@ -23,6 +25,7 @@ def eval_dialog(
 ) -> DialogEval:
     try:
         completion = client.chat_completion(
+            model="meta-llama/Llama-3.3-70B-Instruct",
             messages=[
                 {
                     "role": "user",
@@ -31,9 +34,10 @@ def eval_dialog(
             ],
             response_format={"type": "json", "value": DialogEval.model_json_schema()},
         )
-    except HTTPError:
+    except HTTPError as e:
         return DialogEval(
-            result="Error", reason="Error while interacting with the Hugging Face API."
+            result="Error",
+            reason=f"Error while interacting with the Hugging Face API. {e}",
         )
 
     message = completion.choices[0].message
@@ -48,9 +52,15 @@ def eval_dialog(
 # Save to HF as Parquet. Dataset can be previewed here:
 # https://huggingface.co/datasets/dvcorg/test-datachain-llm-eval/viewer
 (
-    dc.read_csv("hf://datasets/infinite-dataset-hub/MobilePlanAssistant/data.csv")
-    .settings(parallel=10)
-    .setup(client=lambda: InferenceClient("meta-llama/Llama-3.1-70B-Instruct"))
+    dc.read_csv(
+        "hf://datasets/infinite-dataset-hub/MobilePlanAssistant/data.csv", source=False
+    )
+    .settings(parallel=True)
+    .setup(
+        client=lambda: InferenceClient(
+            provider="hf-inference", api_key=os.environ["HF_TOKEN"]
+        )
+    )
     .map(response=eval_dialog)
     .to_parquet("hf://datasets/dvcorg/test-datachain-llm-eval/data.parquet")
 )

{datachain-0.20.1 → datachain-0.20.2}/src/datachain/client/hf.py

@@ -21,6 +21,9 @@ def _wrap_class(sync_fs_class):
     asynchronous to False by default. This is similar to other Async FS
     we initialize. E.g. it means we don't break things in Jupyter where code
     run in async.
+
+    This also fixes write operations by ensuring they are properly forwarded
+    to the underlying filesystem without async buffering issues.
     """
     from fsspec.implementations.asyn_wrapper import AsyncFileSystemWrapper
 
@@ -29,6 +32,13 @@ def _wrap_class(sync_fs_class):
             sync_fs = sync_fs_class(*args, **kwargs)
             super().__init__(sync_fs, asynchronous=False)
 
+        def open(self, path, mode="rb", **kwargs):
+            # Override open to ensure write operations work correctly.
+            # It seems to be a bug in the fsspec wrapper. It avoids
+            # wrapping open() explicitly but also doesn't redirect it to
+            # sync filesystem.
+            return self.sync_fs.open(path, mode, **kwargs)
+
     GeneratedAsyncFileSystemWrapper.__name__ = f"Async{sync_fs_class.__name__}Wrapper"
     return GeneratedAsyncFileSystemWrapper
 

{datachain-0.20.1 → datachain-0.20.2}/src/datachain.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: datachain
-Version: 0.20.1
+Version: 0.20.2
 Summary: Wrangle unstructured AI data at scale
 Author-email: Dmitry Petrov <support@dvc.org>
 License-Expression: Apache-2.0