DeepFabric 4.8.0__py3-none-any.whl → 4.8.2__py3-none-any.whl

deepfabric/evaluation/backends/transformers_backend.py

@@ -2,13 +2,13 @@ import json
 import logging
 import sys
 
+from functools import cached_property
 from typing import Any
 
-import torch
-
 from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
 
 from ...schemas import ToolDefinition
+from ...utils import import_optional_dependency
 from ..inference import InferenceBackend, InferenceConfig, ModelResponse
 from .tool_call_parsers import ToolCallParser, get_parser
 
@@ -29,6 +29,30 @@ logger = logging.getLogger(__name__)
 class TransformersBackend(InferenceBackend):
     """Inference backend using HuggingFace Transformers."""
 
+    @cached_property
+    def _torch(self) -> Any:
+        """Dynamically import 'torch' and verify its availability.
+
+        Returns:
+            The imported torch module.
+
+        Raises:
+            ModuleNotFoundError: If 'torch' is not installed in the environment.
+        """
+        return import_optional_dependency("torch", "training")
+
+    @cached_property
+    def _peft(self) -> Any:
+        """Dynamically import 'peft' and verify its availability.
+
+        Returns:
+            The imported peft module.
+
+        Raises:
+            ModuleNotFoundError: If 'peft' is not installed in the environment.
+        """
+        return import_optional_dependency("peft", "training")
+
     def __init__(self, config: InferenceConfig):
         """Initialize Transformers backend.
 
@@ -47,22 +71,22 @@ class TransformersBackend(InferenceBackend):
             # Get device from pre-loaded model
             self.device = str(next(config.model.parameters()).device)
         # Auto-detect best available device
-        elif torch.cuda.is_available():
+        elif self._torch.cuda.is_available():
             self.device = "cuda"
-        elif torch.backends.mps.is_available():
+        elif self._torch.backends.mps.is_available():
             self.device = "mps"
         else:
             self.device = "cpu"
 
         # Determine dtype based on device
         if self.device == "cuda" or self.device.startswith("cuda:"):
-            dtype = torch.float16
+            dtype = self._torch.float16
             device_map = "auto"
         elif self.device == "mps":
-            dtype = torch.float32  # MPS works best with float32
+            dtype = self._torch.float32  # MPS works best with float32
             device_map = None
         else:
-            dtype = torch.float32
+            dtype = self._torch.float32
             device_map = None
 
         # Handle pre-loaded model case - skip all loading logic
@@ -138,9 +162,9 @@ class TransformersBackend(InferenceBackend):
                         load_in_4bit=config.load_in_4bit,
                     )
                     # Load LoRA adapter using PEFT
-                    from peft import PeftModel  # noqa: PLC0415
-
-                    self.model = PeftModel.from_pretrained(self.model, config.adapter_path)
+                    self.model = self._peft.PeftModel.from_pretrained(
+                        self.model, config.adapter_path
+                    )
                 else:
                     # Load merged model or base model directly
                     self.model, self.tokenizer = FastLanguageModel.from_pretrained(
@@ -172,9 +196,7 @@ class TransformersBackend(InferenceBackend):
 
             # Load PEFT adapter if provided
             if config.adapter_path:
-                from peft import PeftModel  # noqa: PLC0415
-
-                self.model = PeftModel.from_pretrained(self.model, config.adapter_path)
+                self.model = self._peft.PeftModel.from_pretrained(self.model, config.adapter_path)
 
             # Move to device if not using device_map
             if self.device in ("cpu", "mps"):
@@ -217,7 +239,7 @@ class TransformersBackend(InferenceBackend):
         ).to(self.model.device)
 
         # Generate with optimizations
-        with torch.no_grad():
+        with self._torch.no_grad():
             outputs = self.model.generate(
                 **inputs,
                 max_new_tokens=self.config.max_tokens,
@@ -273,7 +295,7 @@ class TransformersBackend(InferenceBackend):
         ).to(self.model.device)
 
         # Generate batch with optimizations
-        with torch.no_grad():
+        with self._torch.no_grad():
             outputs = self.model.generate(
                 **inputs,
                 max_new_tokens=self.config.max_tokens,
@@ -316,8 +338,8 @@ class TransformersBackend(InferenceBackend):
             del self.model
         if hasattr(self, "tokenizer"):
             del self.tokenizer
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
+        if self._torch.cuda.is_available():
+            self._torch.cuda.empty_cache()
 
     def _format_prompt(
         self,

deepfabric/evaluation/evaluator.py

@@ -24,6 +24,25 @@ from .reporters import BaseReporter, CloudReporter, FileReporter, MultiReporter
 
 console = Console()
 
+# Mapping for legacy conversation_type values
+_CONVERSATION_TYPE_ALIASES = {
+    "chain_of_thought": "cot",
+}
+
+
+def _normalize_conversation_type(value: str) -> str:
+    """Normalize conversation_type to valid values.
+
+    Handles legacy values like 'chain_of_thought' -> 'cot'.
+
+    Args:
+        value: Raw conversation_type value from dataset
+
+    Returns:
+        Normalized value ('basic' or 'cot')
+    """
+    return _CONVERSATION_TYPE_ALIASES.get(value, value)
+
 
 class EvaluatorConfig(BaseModel):
     """Configuration for evaluation run."""
@@ -247,9 +266,10 @@ class Evaluator:
         # Convert sample dict to Conversation object
         conversation = Conversation.model_validate(sample)
 
-        # Determine conversation type from metadata
+        # Determine conversation type from metadata (normalize legacy values)
         metadata = conversation.metadata or {}
-        conv_type = metadata.get("conversation_type", "basic")
+        raw_conv_type = metadata.get("conversation_type", "basic")
+        conv_type = _normalize_conversation_type(raw_conv_type)
         reasoning_style = metadata.get("reasoning_style")
         agent_mode = metadata.get("agent_mode")
 

deepfabric/schemas.py

@@ -136,7 +136,7 @@ class MCPInputSchemaProperty(BaseModel):
 
     model_config = {"extra": "allow"}
 
-    type: str = Field(default="string", description="JSON Schema type")
+    type: str | list[str] = Field(default="string", description="JSON Schema type (string or array for nullable)")
     description: str = Field(default="", description="Property description")
     default: Any | None = Field(default=None, description="Default value")
 
@@ -159,7 +159,7 @@ class MCPToolDefinition(BaseModel):
     See: https://modelcontextprotocol.io/specification/2025-06-18/schema#tool
     """
 
-    model_config = {"extra": "allow"}
+    model_config = {"extra": "allow", "populate_by_name": True}
 
     name: str = Field(description="Tool name")
     description: str = Field(default="", description="Tool description")
@@ -367,7 +367,15 @@ class ToolDefinition(BaseModel):
         required_params = set(input_schema.required)
 
         for param_name, param_props in input_schema.properties.items():
-            df_type = type_mapping.get(param_props.type, "str")
+            # Handle type as either string or array (for nullable types like ["string", "null"])
+            param_type = param_props.type
+            if isinstance(param_type, list):
+                # Extract the primary type (non-null type from array)
+                primary_type = next((t for t in param_type if t != "null"), "string")
+            else:
+                primary_type = param_type
+
+            df_type = type_mapping.get(primary_type, "str")
             default_str = str(param_props.default) if param_props.default is not None else ""
 
             parameters.append(

deepfabric/training/callback.py

@@ -298,9 +298,10 @@ class DeepFabricCallback:
         Returns:
             Model name or None
         """
-        # Try args first
-        if hasattr(args, "model_name_or_path"):
-            return args.model_name_or_path
+        # Try args first (model_name_or_path exists on SFTConfig and similar subclasses)
+        model_name_or_path = getattr(args, "model_name_or_path", None)
+        if model_name_or_path is not None:
+            return model_name_or_path
 
         # Try model config
         if model is not None:
@@ -310,8 +311,9 @@ class DeepFabricCallback:
                 return model.name_or_path
 
         # Try output_dir as fallback
-        if hasattr(args, "output_dir"):
-            return os.path.basename(args.output_dir)
+        output_dir = getattr(args, "output_dir", None)
+        if output_dir is not None:
+            return os.path.basename(output_dir)
 
         return None
 

deepfabric/utils.py

@@ -1,9 +1,12 @@
 import ast
 import asyncio
+import importlib
 import json
 import os
 import re
 
+from typing import Any
+
 VALIDATION_ERROR_INDICATORS = [
     "validation error",
     "value error",
@@ -162,3 +165,33 @@ def get_bool_env(key: str, default: bool = False) -> bool:
     if val is None:
         return default
     return val.lower() in ("1", "true", "yes", "on")
+
+
+def import_optional_dependency(
+    module_name: str,
+    extra: str | None = None,
+) -> Any:
+    """
+    Import an optional dependency at runtime.
+
+    Args:
+        module_name (str): The name of the module to import.
+        extra (str | None): The optional dependency group providing this module.
+
+    Returns:
+        Any: The imported module.
+
+    Raises:
+        ModuleNotFoundError: If the module is not installed.
+    """
+    try:
+        return importlib.import_module(module_name)
+    except ModuleNotFoundError:
+        if extra:
+            msg = (
+                f"The '{module_name}' library is required for the '{extra}' features. "
+                f"Please install it using: pip install 'deepfabric[{extra}]'"
+            )
+        else:
+            msg = f"The '{module_name}' library is required but is not installed."
+        raise ModuleNotFoundError(msg) from None

{deepfabric-4.8.0.dist-info → deepfabric-4.8.2.dist-info}/METADATA

@@ -1,11 +1,10 @@
 Metadata-Version: 2.4
 Name: DeepFabric
-Version: 4.8.0
+Version: 4.8.2
 Summary: Curate High Quality Datasets, Train, Evaluate and Ship
 Author-email: DeepFabric Team <oss@alwaysfurther.ai>
 License-File: LICENSE
 Requires-Python: >=3.10
-Requires-Dist: accelerate>=0.20.0
 Requires-Dist: anthropic>=0.75.0
 Requires-Dist: click>=8.1.7
 Requires-Dist: componentize-py>=0.19.3
@@ -19,7 +18,6 @@ Requires-Dist: ollama>=0.6.1
 Requires-Dist: openai>=1.107.2
 Requires-Dist: outlines==1.2.9
 Requires-Dist: packaging>=25.0
-Requires-Dist: peft>=0.7.0
 Requires-Dist: posthog>=3.0.0
 Requires-Dist: protobuf>=3.20.0
 Requires-Dist: pydantic>=2.0.0
@@ -27,9 +25,7 @@ Requires-Dist: pyyaml>=6.0.1
 Requires-Dist: rich>=13.0.0
 Requires-Dist: sentencepiece>=0.1.99
 Requires-Dist: spin-sdk>=3.4.1
-Requires-Dist: torch>=2.4.0
 Requires-Dist: transformers>=4.57.1
-Requires-Dist: trl>=0.26.2
 Provides-Extra: dev
 Requires-Dist: bandit>=1.7.10; extra == 'dev'
 Requires-Dist: mermaid-py>=0.2.0; extra == 'dev'
@@ -42,6 +38,11 @@ Requires-Dist: ruff>=0.1.0; extra == 'dev'
 Provides-Extra: docs
 Requires-Dist: mkdocs-material>=9.0.0; extra == 'docs'
 Requires-Dist: mkdocstrings[python]>=0.30.0; extra == 'docs'
+Provides-Extra: training
+Requires-Dist: accelerate>=0.20.0; extra == 'training'
+Requires-Dist: peft>=0.7.0; extra == 'training'
+Requires-Dist: torch>=2.4.0; extra == 'training'
+Requires-Dist: trl>=0.26.2; extra == 'training'
 Description-Content-Type: text/markdown
 
 <div align="center">

{deepfabric-4.8.0.dist-info → deepfabric-4.8.2.dist-info}/RECORD

@@ -21,17 +21,17 @@ deepfabric/loader.py,sha256=YNTGZZE-POjR0BIlx6WCT4bIzf0T4lW_fQl7ev9UFqE,18584
 deepfabric/metrics.py,sha256=iwtNHBX4ZTYUg2FZgtFcG3U0e9RlV2c1cm1Kp34FeWU,6129
 deepfabric/progress.py,sha256=3XQQrf2pUZlyd-8eRcNATH1v0Oi8JMedVHGbhPcca-8,9354
 deepfabric/prompts.py,sha256=JVFMeeBa2qqOMvmP_xx8bWzZ6ot9eyqOP3u8XzzPx3g,10290
-deepfabric/schemas.py,sha256=ckzAjDc6IlC8Y-Pi2hyYRqcSwARX7z_GELuCypXuSgI,37401
+deepfabric/schemas.py,sha256=r8qQuu19o9ev0JNxcChjbpWym9m5WLItyxw8szYLSjI,37867
 deepfabric/stream_simulator.py,sha256=GzvAxWxHVsuTwgXlqwXNfrTUDn6sND2kJOoQuYg88FA,3028
 deepfabric/topic_manager.py,sha256=6YxMO6dQHaGyxghsI8iNJGP1miaekBe5Mh1WdYeLqdI,11164
 deepfabric/topic_model.py,sha256=i_wYpw2kUl8NLodOSaqNu-C4_d6caYT1kPe_vkKjoyw,707
 deepfabric/tree.py,sha256=Kxl2iLHU55xPq2MwdoLM0-M2nZRx51bRj9FM36jqs-M,14933
 deepfabric/tui.py,sha256=9ETtGFQk26U9PQ2b5foplVYDKxaFGd-8UqK7uSKyHwE,50480
 deepfabric/update_checker.py,sha256=AUa9iUdkGNzu7tWkQRxIlF19YRmKLetwxu-Ys2ONS8Y,5145
-deepfabric/utils.py,sha256=ve6tku_-jgW_ZIkh9osUEQ3C_03J6R_zOw0Xf5UGJYc,4891
+deepfabric/utils.py,sha256=a9G6VTw52UdddTFoMw-JjunjawtPN54N275-XGPL2cQ,5822
 deepfabric/validation.py,sha256=1x1X_45kyI0w_FCdUiNdvy4LQu3B0KVR-fyvLkrKEGw,5125
 deepfabric/evaluation/__init__.py,sha256=7xMLmYXaNC1U7qf88S9fMxWTABoDRiOcimSYfCt_PSo,1224
-deepfabric/evaluation/evaluator.py,sha256=ExUrL5Zil4DzibzjzngA7dfxnmGIVp9H7319FhLHYmk,33918
+deepfabric/evaluation/evaluator.py,sha256=qNowle5v2ukDJ11igNOCParlBfXT8QUeOvXx6sSJ_Ug,34480
 deepfabric/evaluation/inference.py,sha256=y7JA0IsBDwe0sJzVQeItYHAV5wUJn6Bjp1Wsp3r7qYQ,7644
 deepfabric/evaluation/metrics.py,sha256=ITNevYj7CBXzYs-rYhsihO6-rE9n30CYRaVUfdTbcFQ,12026
 deepfabric/evaluation/parser.py,sha256=AXyiCtNV4rueZQxLE_GqqkFNeDAewGoC--0vXHW-jW8,10603
@@ -39,7 +39,7 @@ deepfabric/evaluation/backends/__init__.py,sha256=GqC0FfpWmtgJmjHd0kVKNg7g-NjhRo
 deepfabric/evaluation/backends/llm_eval_backend.py,sha256=4jp5tnTp7v_0pHCGhcPbI55ig79-eVxdzooesi2PymA,18827
 deepfabric/evaluation/backends/ollama_backend.py,sha256=mtPp1JtIDRjb76X_rTa1jS1ETzMjte8t3WJjuYV1oDQ,4372
 deepfabric/evaluation/backends/tool_call_parsers.py,sha256=Ufg4Xt3mrDS-WbGor6tOOr4xZNCHk3Co2C-z_o-pAkM,14126
-deepfabric/evaluation/backends/transformers_backend.py,sha256=WcqB9gkayQpjx2Em00lhzJg8RcWdQEYbctDNXLzFChA,14484
+deepfabric/evaluation/backends/transformers_backend.py,sha256=f3rbFxjWdv2NhDvlMfl0YwFUkfh0i5dlM3JKYeoJgvQ,15243
 deepfabric/evaluation/evaluators/__init__.py,sha256=NdH65YvanskRGe6r7JepkTNGGt8xA-GLugagU3VQ_WM,353
 deepfabric/evaluation/evaluators/base.py,sha256=1TiLr-_oF9dRmdSgJs94dDbf0gTwRS8TGGz2C1Z3nag,2946
 deepfabric/evaluation/evaluators/registry.py,sha256=VGeb1AHFGkn9TLpcqfuGIZi1jgh7Qw0NNILT6z3Se6M,2171
@@ -66,11 +66,11 @@ deepfabric/tools/loader.py,sha256=Bv56D-76JChlK_QXfHLw_rneGLZYRhkn5ETbJMIdJsA,29
 deepfabric/tools/mcp_client.py,sha256=uQRrlDSVwF0ZatOl9bidBNU7IgXgJKQU-xG50dK0Uy4,23377
 deepfabric/training/__init__.py,sha256=MJazTELfrTB15rIiCE04hDeUL8LSSg4-4LWWG6j2BRw,1566
 deepfabric/training/api_key_prompt.py,sha256=pSIMX3eDGyV9x_r7MHE4TyIsIB2SqYb8gKCdAtTY-q8,9371
-deepfabric/training/callback.py,sha256=NUNrMAEYKt9kPjrX9mckvs8H4uoeVtRWPsfrjW90fWI,13051
+deepfabric/training/callback.py,sha256=5zdifbHA2PWILHl2cVFyO65aW7cGAQhcvDqm3s8_I0Q,13221
 deepfabric/training/dataset_utils.py,sha256=klx8DoawEwuMigBDP-RpMAfe7FvYxRbhj599MErxBr4,7313
 deepfabric/training/metrics_sender.py,sha256=ZCyvMv5hRu8XJnQYVGXJ9wh7HEMJ0l3Ktyi8_etOpZs,10833
-deepfabric-4.8.0.dist-info/METADATA,sha256=N3ttE4MjkF2rk_Zq75-wjDds3BM0L8jEciHz8wnAnx0,20427
-deepfabric-4.8.0.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-deepfabric-4.8.0.dist-info/entry_points.txt,sha256=zatevils13hfs8x29_vmUyivQ6rTtq7hE2RBusZw1Fo,50
-deepfabric-4.8.0.dist-info/licenses/LICENSE,sha256=-qRt8wmrhQ9aMf7KhmZXc2vrTETYZF-6_T1KCeUhvHY,11340
-deepfabric-4.8.0.dist-info/RECORD,,
+deepfabric-4.8.2.dist-info/METADATA,sha256=15ZBOITSr6pigZYD-GBn_IwRlnzftZ64hwrXldKr1mg,20536
+deepfabric-4.8.2.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+deepfabric-4.8.2.dist-info/entry_points.txt,sha256=zatevils13hfs8x29_vmUyivQ6rTtq7hE2RBusZw1Fo,50
+deepfabric-4.8.2.dist-info/licenses/LICENSE,sha256=-qRt8wmrhQ9aMf7KhmZXc2vrTETYZF-6_T1KCeUhvHY,11340
+deepfabric-4.8.2.dist-info/RECORD,,