vllm-judge 0.1.1__tar.gz → 0.1.3__tar.gz

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.1
 Name: vllm_judge
-Version: 0.1.1
+Version: 0.1.3
 Summary: LLM-as-a-Judge evaluations for vLLM hosted models
 Author: TrustyAI team
 Author-email: Sai Chandra Pandraju <saichandrapandraju@gmail.com>
@@ -37,18 +37,21 @@ Requires-Dist: isort>=5.12.0; extra == "dev"
 Requires-Dist: flake8>=6.0.0; extra == "dev"
 Requires-Dist: mypy>=1.0.0; extra == "dev"
 
+[![PyPI version](https://img.shields.io/pypi/v/vllm-judge.svg)
+](https://pypi.org/project/vllm-judge/)
+
 # vLLM Judge
 
-A lightweight library for LLM-as-a-Judge evaluations using vLLM hosted models. Please refer the [documentation](https://saichandrapandraju.github.io/vllm_judge/) for usage details.
+A lightweight library for LLM-as-a-Judge evaluations using vLLM hosted models. Evaluate LLM inputs & outputs at scale with just a few lines of code. From simple scoring to complex safety checks, vLLM Judge adapts to your needs. Please refer the [documentation](https://saichandrapandraju.github.io/vllm_judge/) for usage details.
 
 ## Features
 
 - 🚀 **Simple Interface**: Single `evaluate()` method that adapts to any use case
 - 🎯 **Pre-built Metrics**: 20+ ready-to-use evaluation metrics
+- 🛡️ **Model-Specific Support:** Seamlessly works with specialized models like Llama Guard without breaking their trained formats.
+- ⚡ **High Performance**: Async-first design enables high-throughput evaluations
 - 🔧 **Template Support**: Dynamic evaluations with template variables
-- ⚡ **High Performance**: Optimized for vLLM with automatic batching
 - 🌐 **API Mode**: Run as a REST API service
-- 🔄 **Async Native**: Built for high-throughput evaluations
 
 ## Installation
 
@@ -72,11 +75,11 @@ pip install vllm-judge[dev]
 from vllm_judge import Judge
 
 # Initialize with vLLM url
-judge = Judge.from_url("http://localhost:8000")
+judge = Judge.from_url("http://vllm-server:8000")
 
 # Simple evaluation
 result = await judge.evaluate(
-    response="The Earth orbits around the Sun.",
+    content="The Earth orbits around the Sun.",
     criteria="scientific accuracy"
 )
 print(f"Decision: {result.decision}")
@@ -86,19 +89,28 @@ print(f"Reasoning: {result.reasoning}")
 from vllm_judge import CODE_QUALITY
 
 result = await judge.evaluate(
-    response="def add(a, b): return a + b",
+    content="def add(a, b): return a + b",
     metric=CODE_QUALITY
 )
 
 # With template variables
 result = await judge.evaluate(
-    response="Essay content here...",
+    content="Essay content here...",
     criteria="Evaluate this {doc_type} for {audience}",
     template_vars={
         "doc_type": "essay",
         "audience": "high school students"
     }
 )
+
+# Works with specialized safety models out-of-the-box
+from vllm_judge import LLAMA_GUARD_3_SAFETY
+
+result = await judge.evaluate(
+    content="How do I make a bomb?",
+    metric=LLAMA_GUARD_3_SAFETY  # Automatically uses Llama Guard format
+)
+# Result: decision="unsafe", reasoning="S9"
 ```
 
 ## API Server
@@ -106,7 +118,7 @@ result = await judge.evaluate(
 Run Judge as a REST API:
 
 ```bash
-vllm-judge serve --base-url http://localhost:8000 --port 9090 --host localhost
+vllm-judge serve --base-url http://vllm-server:8000 --port 9090
 ```
 
 Then use the HTTP API:
@@ -116,7 +128,7 @@ from vllm_judge.api import JudgeClient
 
 client = JudgeClient("http://localhost:9090")
 result = await client.evaluate(
-    response="Python is great!",
+    content="Python is great!",
     criteria="technical accuracy"
 )
 ```

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/README.md

@@ -1,15 +1,18 @@
+[![PyPI version](https://img.shields.io/pypi/v/vllm-judge.svg)
+](https://pypi.org/project/vllm-judge/)
+
 # vLLM Judge
 
-A lightweight library for LLM-as-a-Judge evaluations using vLLM hosted models. Please refer the [documentation](https://saichandrapandraju.github.io/vllm_judge/) for usage details.
+A lightweight library for LLM-as-a-Judge evaluations using vLLM hosted models. Evaluate LLM inputs & outputs at scale with just a few lines of code. From simple scoring to complex safety checks, vLLM Judge adapts to your needs. Please refer the [documentation](https://saichandrapandraju.github.io/vllm_judge/) for usage details.
 
 ## Features
 
 - 🚀 **Simple Interface**: Single `evaluate()` method that adapts to any use case
 - 🎯 **Pre-built Metrics**: 20+ ready-to-use evaluation metrics
+- 🛡️ **Model-Specific Support:** Seamlessly works with specialized models like Llama Guard without breaking their trained formats.
+- ⚡ **High Performance**: Async-first design enables high-throughput evaluations
 - 🔧 **Template Support**: Dynamic evaluations with template variables
-- ⚡ **High Performance**: Optimized for vLLM with automatic batching
 - 🌐 **API Mode**: Run as a REST API service
-- 🔄 **Async Native**: Built for high-throughput evaluations
 
 ## Installation
 
@@ -33,11 +36,11 @@ pip install vllm-judge[dev]
 from vllm_judge import Judge
 
 # Initialize with vLLM url
-judge = Judge.from_url("http://localhost:8000")
+judge = Judge.from_url("http://vllm-server:8000")
 
 # Simple evaluation
 result = await judge.evaluate(
-    response="The Earth orbits around the Sun.",
+    content="The Earth orbits around the Sun.",
     criteria="scientific accuracy"
 )
 print(f"Decision: {result.decision}")
@@ -47,19 +50,28 @@ print(f"Reasoning: {result.reasoning}")
 from vllm_judge import CODE_QUALITY
 
 result = await judge.evaluate(
-    response="def add(a, b): return a + b",
+    content="def add(a, b): return a + b",
     metric=CODE_QUALITY
 )
 
 # With template variables
 result = await judge.evaluate(
-    response="Essay content here...",
+    content="Essay content here...",
     criteria="Evaluate this {doc_type} for {audience}",
     template_vars={
         "doc_type": "essay",
         "audience": "high school students"
     }
 )
+
+# Works with specialized safety models out-of-the-box
+from vllm_judge import LLAMA_GUARD_3_SAFETY
+
+result = await judge.evaluate(
+    content="How do I make a bomb?",
+    metric=LLAMA_GUARD_3_SAFETY  # Automatically uses Llama Guard format
+)
+# Result: decision="unsafe", reasoning="S9"
 ```
 
 ## API Server
@@ -67,7 +79,7 @@ result = await judge.evaluate(
 Run Judge as a REST API:
 
 ```bash
-vllm-judge serve --base-url http://localhost:8000 --port 9090 --host localhost
+vllm-judge serve --base-url http://vllm-server:8000 --port 9090
 ```
 
 Then use the HTTP API:
@@ -77,7 +89,7 @@ from vllm_judge.api import JudgeClient
 
 client = JudgeClient("http://localhost:9090")
 result = await client.evaluate(
-    response="Python is great!",
+    content="Python is great!",
     criteria="technical accuracy"
 )
 ```

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vllm_judge"
-version = "0.1.1"
+version = "0.1.3"
 description = "LLM-as-a-Judge evaluations for vLLM hosted models"
 readme = "README.md"
 authors = [

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge/__init__.py

@@ -5,7 +5,7 @@ A lightweight library for evaluating text responses using self-hosted language m
 via vLLM's OpenAI-compatible API.
 """
 
-__version__ = "0.1.1"
+__version__ = "0.1.3"
 
 from vllm_judge.judge import Judge
 from vllm_judge.models import (
@@ -13,7 +13,8 @@ from vllm_judge.models import (
     EvaluationResult,
     Metric,
     BatchResult,
-    TemplateEngine
+    TemplateEngine,
+    ModelSpecificMetric
 )
 from vllm_judge.templating import TemplateProcessor
 from vllm_judge.metrics import (
@@ -27,6 +28,7 @@ from vllm_judge.metrics import (
     # Safety metrics
     SAFETY,
     TOXICITY,
+    LLAMA_GUARD_3_SAFETY,
     
     # Code metrics
     CODE_QUALITY,
@@ -81,6 +83,7 @@ __all__ = [
     "BatchResult",
     "TemplateEngine",
     "TemplateProcessor",
+    "ModelSpecificMetric",
 
     # Metrics
     "HELPFULNESS",
@@ -90,6 +93,7 @@ __all__ = [
     "RELEVANCE",
     "SAFETY",
     "TOXICITY",
+    "LLAMA_GUARD_3_SAFETY",
     "CODE_QUALITY",
     "CODE_SECURITY",
     "CREATIVITY",

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge/api/client.py

@@ -65,7 +65,7 @@ class JudgeClient:
     
     async def evaluate(
         self,
-        response: Union[str, Dict[str, str]],
+        content: Union[str, Dict[str, str]],
         criteria: str = None,
         rubric: Union[str, Dict[Union[int, float], str]] = None,
         scale: Optional[Tuple[int, int]] = None,
@@ -87,7 +87,7 @@ class JudgeClient:
             EvaluationResult
         """
         request = EvaluateRequest(
-            response=response,
+            response=content,
             criteria=criteria,
             rubric=rubric,
             scale=list(scale) if scale else None,

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge/cli.py

@@ -75,7 +75,7 @@ def evaluate(
             # Use API client
             async with JudgeClient(api_url) as client:
                 result = await client.evaluate(
-                    response=response,
+                    content=response,
                     criteria=criteria,
                     metric=metric,
                     scale=scale,
@@ -91,7 +91,7 @@ def evaluate(
             judge = Judge.from_url(base_url, model=model)
             async with judge:
                 result = await judge.evaluate(
-                    response=response,
+                    content=response,
                     criteria=criteria,
                     metric=metric,
                     scale=scale,

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge/judge.py

@@ -2,7 +2,7 @@ import json
 import re
 from typing import Union, Dict, List, Optional, Tuple, Any, Callable
 
-from vllm_judge.models import JudgeConfig, EvaluationResult, Metric, BatchResult, TemplateEngine
+from vllm_judge.models import JudgeConfig, EvaluationResult, Metric, BatchResult, TemplateEngine, ModelSpecificMetric
 from vllm_judge.client import VLLMClient
 from vllm_judge.prompts import PromptBuilder
 from vllm_judge.batch import BatchProcessor
@@ -14,6 +14,9 @@ from vllm_judge.exceptions import (
     MetricNotFoundError,
     VLLMJudgeError
 )
+import logging
+
+logger = logging.getLogger(__name__)
 
 
 class Judge:
@@ -60,7 +63,7 @@ class Judge:
     
     async def evaluate(
         self,
-        response: Union[str, Dict[str, str]],
+        content: Union[str, Dict[str, str]],
         criteria: str = None,
         rubric: Union[str, Dict[Union[int, float], str]] = None,
         scale: Optional[Tuple[int, int]] = None,
@@ -76,7 +79,7 @@ class Judge:
         Universal evaluation method that adapts to use case.
         
         Args:
-            response: String for single evaluation, dict {"a": ..., "b": ...} for comparison
+            content: String for single evaluation, dict {"a": ..., "b": ...} for comparison
             criteria: What to evaluate for (can contain template variables)
             rubric: Instructions for evaluation, can be string or dict containing mapping of score to description (can contain template variables)
             scale: Optional numeric scale (min, max)
@@ -96,6 +99,22 @@ class Judge:
             MetricNotFoundError: If metric name not found
             ParseError: If unable to parse model response
         """
+        # Handle model-specific metrics
+        if isinstance(metric, ModelSpecificMetric):
+            assert isinstance(content, str), "Model-specific metrics only support string content for now"
+
+            # logger.info(f"Evaluating model-specific metric {metric.name}.")
+            logger.info(f"We assume you're using {metric.model_pattern} type model. If not, please do not use this metric and use a normal metric instead.")
+            # Skip ALL our formatting
+            messages = [{"role": "user", "content": content}]
+            
+            # vLLM applies model's chat template automatically
+            llm_response = await self._call_model(messages)
+            
+            # Use metric's parser
+            return metric.parser_func(llm_response)
+        
+        # Handle normal metrics
         # Handle metric parameter
         metric_template_vars = {}
         
@@ -138,7 +157,7 @@ class Judge:
         
         # Build messages
         messages = PromptBuilder.build_messages(
-            response=response,
+            response=content,
             criteria=criteria,
             rubric=rubric,
             scale=scale,
@@ -149,14 +168,7 @@ class Judge:
         )
         
         # Get LLM response
-        try:
-            if self.config.use_chat_api:
-                llm_response = await self.client.chat_completion(messages)
-            else:
-                prompt = PromptBuilder.format_messages_as_text(messages)
-                llm_response = await self.client.completion(prompt)
-        except Exception as e:
-            raise VLLMJudgeError(f"Failed to get model response: {e}")
+        llm_response = await self._call_model(messages)
         
         # Parse response
         result = self._parse_response(llm_response)
@@ -168,6 +180,21 @@ class Judge:
         
         return result
     
+    async def _call_model(self, messages: List[Dict[str, str]]) -> str:
+        """
+        Call the model with the given messages.
+        """
+        try:
+            if self.config.use_chat_api:
+                llm_response = await self.client.chat_completion(messages)
+            else:
+                prompt = PromptBuilder.format_messages_as_text(messages)
+                llm_response = await self.client.completion(prompt)
+            return llm_response
+        except Exception as e:
+            raise VLLMJudgeError(f"Failed to get model response: {e}")
+
+    
     def _parse_response(self, response: str) -> EvaluationResult:
         """
         Parse LLM response into EvaluationResult.

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge/metrics.py

@@ -1,5 +1,6 @@
 from typing import Dict
-from vllm_judge.models import Metric,TemplateEngine
+from vllm_judge.models import Metric, TemplateEngine, ModelSpecificMetric
+from vllm_judge.utils import parse_llama_guard_3
 
 # Registry for built-in metrics
 BUILTIN_METRICS: Dict[str, Metric] = {}
@@ -11,6 +12,13 @@ def create_builtin_metric(metric: Metric) -> Metric:
     return metric
 
 
+# Llama Guard 3 safety metric
+LLAMA_GUARD_3_SAFETY = create_builtin_metric(ModelSpecificMetric(
+    name="llama_guard_3_safety",
+    model_pattern="llama_guard_3",
+    parser_func=parse_llama_guard_3
+))
+
 # General purpose metrics
 HELPFULNESS = create_builtin_metric(Metric(
     name="helpfulness",

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge/models.py

@@ -1,4 +1,4 @@
-from typing import Optional, Any, Dict, Union, List, Tuple
+from typing import Optional, Any, Dict, Union, List, Tuple, Callable
 from pydantic import BaseModel, Field, field_validator, ConfigDict
 from enum import Enum
 
@@ -159,6 +159,15 @@ class Metric:
     def __repr__(self):
         return f"Metric(name='{self.name}', criteria='{self.criteria}', template_engine='{self.template_engine}')"
 
+# Base class for model-specific metrics
+class ModelSpecificMetric(Metric):
+    """Metric that bypasses our prompt formatting."""
+    
+    def __init__(self, name: str, model_pattern: str, parser_func: Callable[[str], EvaluationResult]):
+        super().__init__(name=name, criteria="model-specific evaluation")
+        self.model_pattern = model_pattern
+        self.parser_func = parser_func
+        # self.is_model_specific = True  # Flag for special handling
 
 class BatchResult(BaseModel):
     """Result of batch evaluation."""

vllm_judge-0.1.3/src/vllm_judge/utils.py

@@ -0,0 +1,14 @@
+from vllm_judge.models import EvaluationResult
+
+# Llama Guard 3 parser
+def parse_llama_guard_3(response: str) -> EvaluationResult:
+    """Parse Llama Guard 3's 'safe/unsafe' format."""
+    lines = response.strip().split('\n')
+    is_safe = lines[0].lower().strip() == 'safe'
+    
+    return EvaluationResult(
+        decision="safe" if is_safe else "unsafe",
+        reasoning=lines[1] if len(lines) > 1 else "No violations detected",
+        score=None,
+        metadata={"model_type": "llama_guard_3"}
+    )

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.1
 Name: vllm_judge
-Version: 0.1.1
+Version: 0.1.3
 Summary: LLM-as-a-Judge evaluations for vLLM hosted models
 Author: TrustyAI team
 Author-email: Sai Chandra Pandraju <saichandrapandraju@gmail.com>
@@ -37,18 +37,21 @@ Requires-Dist: isort>=5.12.0; extra == "dev"
 Requires-Dist: flake8>=6.0.0; extra == "dev"
 Requires-Dist: mypy>=1.0.0; extra == "dev"
 
+[![PyPI version](https://img.shields.io/pypi/v/vllm-judge.svg)
+](https://pypi.org/project/vllm-judge/)
+
 # vLLM Judge
 
-A lightweight library for LLM-as-a-Judge evaluations using vLLM hosted models. Please refer the [documentation](https://saichandrapandraju.github.io/vllm_judge/) for usage details.
+A lightweight library for LLM-as-a-Judge evaluations using vLLM hosted models. Evaluate LLM inputs & outputs at scale with just a few lines of code. From simple scoring to complex safety checks, vLLM Judge adapts to your needs. Please refer the [documentation](https://saichandrapandraju.github.io/vllm_judge/) for usage details.
 
 ## Features
 
 - 🚀 **Simple Interface**: Single `evaluate()` method that adapts to any use case
 - 🎯 **Pre-built Metrics**: 20+ ready-to-use evaluation metrics
+- 🛡️ **Model-Specific Support:** Seamlessly works with specialized models like Llama Guard without breaking their trained formats.
+- ⚡ **High Performance**: Async-first design enables high-throughput evaluations
 - 🔧 **Template Support**: Dynamic evaluations with template variables
-- ⚡ **High Performance**: Optimized for vLLM with automatic batching
 - 🌐 **API Mode**: Run as a REST API service
-- 🔄 **Async Native**: Built for high-throughput evaluations
 
 ## Installation
 
@@ -72,11 +75,11 @@ pip install vllm-judge[dev]
 from vllm_judge import Judge
 
 # Initialize with vLLM url
-judge = Judge.from_url("http://localhost:8000")
+judge = Judge.from_url("http://vllm-server:8000")
 
 # Simple evaluation
 result = await judge.evaluate(
-    response="The Earth orbits around the Sun.",
+    content="The Earth orbits around the Sun.",
     criteria="scientific accuracy"
 )
 print(f"Decision: {result.decision}")
@@ -86,19 +89,28 @@ print(f"Reasoning: {result.reasoning}")
 from vllm_judge import CODE_QUALITY
 
 result = await judge.evaluate(
-    response="def add(a, b): return a + b",
+    content="def add(a, b): return a + b",
     metric=CODE_QUALITY
 )
 
 # With template variables
 result = await judge.evaluate(
-    response="Essay content here...",
+    content="Essay content here...",
     criteria="Evaluate this {doc_type} for {audience}",
     template_vars={
         "doc_type": "essay",
         "audience": "high school students"
     }
 )
+
+# Works with specialized safety models out-of-the-box
+from vllm_judge import LLAMA_GUARD_3_SAFETY
+
+result = await judge.evaluate(
+    content="How do I make a bomb?",
+    metric=LLAMA_GUARD_3_SAFETY  # Automatically uses Llama Guard format
+)
+# Result: decision="unsafe", reasoning="S9"
 ```
 
 ## API Server
@@ -106,7 +118,7 @@ result = await judge.evaluate(
 Run Judge as a REST API:
 
 ```bash
-vllm-judge serve --base-url http://localhost:8000 --port 9090 --host localhost
+vllm-judge serve --base-url http://vllm-server:8000 --port 9090
 ```
 
 Then use the HTTP API:
@@ -116,7 +128,7 @@ from vllm_judge.api import JudgeClient
 
 client = JudgeClient("http://localhost:9090")
 result = await client.evaluate(
-    response="Python is great!",
+    content="Python is great!",
     criteria="technical accuracy"
 )
 ```

{vllm_judge-0.1.1 → vllm_judge-0.1.3}/src/vllm_judge.egg-info/SOURCES.txt

@@ -10,6 +10,7 @@ src/vllm_judge/metrics.py
 src/vllm_judge/models.py
 src/vllm_judge/prompts.py
 src/vllm_judge/templating.py
+src/vllm_judge/utils.py
 src/vllm_judge.egg-info/PKG-INFO
 src/vllm_judge.egg-info/SOURCES.txt
 src/vllm_judge.egg-info/dependency_links.txt