File size: 8,255 Bytes
61da702
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
"""
OpenAI SDK Integration Example for HallucinationGuard-Env.

This example demonstrates how to evaluate OpenAI models
(GPT-4, GPT-4o, GPT-3.5) using the HallucinationGuard environment.

Requirements:
    pip install openai requests
"""

import os
from typing import Optional
import requests

# OpenAI SDK
try:
    from openai import OpenAI
except ImportError:
    print("Install OpenAI SDK: pip install openai")
    raise


class HallucinationGuardEvaluator:
    """
    Evaluate OpenAI models for hallucination resistance.

    Features:
    - Supports all OpenAI chat models
    - Handles rate limiting gracefully
    - Tracks calibration and grounding scores
    """

    def __init__(
        self,
        env_base_url: str = "https://samsankar-hallucination-guard-env.hf.space",
        openai_api_key: Optional[str] = None,
        model: str = "gpt-4o-mini"
    ):
        """
        Initialize evaluator.

        Args:
            env_base_url: HallucinationGuard-Env server URL
            openai_api_key: OpenAI API key (or set OPENAI_API_KEY env var)
            model: OpenAI model name
        """
        self.env_base_url = env_base_url.rstrip('/')
        self.model = model
        self.client = OpenAI(api_key=openai_api_key or os.environ.get("OPENAI_API_KEY"))

        # Session for environment
        self.session_id = None
        self.episode_id = None

    def reset_environment(self, difficulty: str = "intermediate") -> dict:
        """
        Start a new evaluation episode.

        Args:
            difficulty: Starting difficulty (beginner, intermediate, advanced)

        Returns:
            Initial observation with question and context
        """
        response = requests.post(
            f"{self.env_base_url}/reset",
            json={"difficulty": difficulty}
        )
        response.raise_for_status()
        data = response.json()

        self.episode_id = data.get("episode_id")
        return data

    def generate_answer(self, question: str, context: str) -> dict:
        """
        Generate an answer using OpenAI model.

        Prompts the model to:
        1. Answer ONLY from the provided context
        2. Provide a confidence score
        3. Cite the source quote

        Args:
            question: The question to answer
            context: The source context

        Returns:
            dict with answer, confidence, source_quote
        """
        prompt = f"""Answer the following question using ONLY the provided context.

IMPORTANT RULES:
1. Answer ONLY from the context - do not use outside knowledge
2. If the answer is not in the context, say "I cannot answer from the provided context"
3. Provide your confidence level (0.0-1.0)
4. Quote the exact passage from the context that supports your answer

CONTEXT:
{context}

QUESTION:
{question}

Respond in JSON format:
{{
    "answer": "your answer here",
    "confidence": 0.85,
    "source_quote": "exact quote from context"
}}

JSON Response:"""

        try:
            response = self.client.chat.completions.create(
                model=self.model,
                messages=[
                    {"role": "system", "content": "You are a precise QA assistant. Always respond in valid JSON format."},
                    {"role": "user", "content": prompt}
                ],
                temperature=0.1,  # Low temperature for factual tasks
                max_tokens=500,
                response_format={"type": "json_object"}
            )

            import json
            content = response.choices[0].message.content
            result = json.loads(content)

            return {
                "answer": result.get("answer", ""),
                "confidence": float(result.get("confidence", 0.5)),
                "source_quote": result.get("source_quote", "")
            }

        except Exception as e:
            print(f"Error generating answer: {e}")
            return {
                "answer": "I cannot answer from the provided context.",
                "confidence": 0.3,
                "source_quote": ""
            }

    def step(self, answer: str, confidence: float, source_quote: str = "") -> dict:
        """
        Submit an answer to the environment.

        Args:
            answer: The answer text
            confidence: Confidence level (0.0-1.0)
            source_quote: Verbatim quote from context

        Returns:
            Observation with reward and feedback
        """
        response = requests.post(
            f"{self.env_base_url}/step",
            json={
                "answer": answer,
                "confidence": confidence,
                "source_quote": source_quote
            }
        )
        response.raise_for_status()
        return response.json()

    def evaluate_episode(
        self,
        num_questions: int = 10,
        difficulty: str = "intermediate"
    ) -> dict:
        """
        Run a complete evaluation episode.

        Args:
            num_questions: Number of questions to evaluate
            difficulty: Starting difficulty level

        Returns:
            Episode statistics
        """
        # Reset environment
        obs = self.reset_environment(difficulty=difficulty)

        total_reward = 0.0
        hallucinations = 0
        correct = 0

        for step_num in range(num_questions):
            # Get current question and context
            question = obs.get("question", "")
            context = obs.get("context", "")

            print(f"\n--- Question {step_num + 1}/{num_questions} ---")
            print(f"Q: {question[:100]}...")

            # Generate answer with OpenAI
            answer_data = self.generate_answer(question, context)
            print(f"A: {answer_data['answer'][:100]}...")
            print(f"Confidence: {answer_data['confidence']:.2f}")

            # Submit to environment
            obs = self.step(
                answer=answer_data["answer"],
                confidence=answer_data["confidence"],
                source_quote=answer_data["source_quote"]
            )

            # Track statistics
            reward = obs.get("reward", 0.0)
            total_reward += reward
            if obs.get("is_hallucination", False):
                hallucinations += 1
            if obs.get("grounding_score", 0) > 0.7:
                correct += 1

            print(f"Reward: {reward:.3f}")
            print(f"Feedback: {obs.get('feedback', '')[:100]}...")

            if obs.get("done", False):
                break

        # Calculate final statistics
        avg_reward = total_reward / max(1, step_num + 1)
        hallucination_rate = hallucinations / max(1, step_num + 1)
        accuracy = correct / max(1, step_num + 1)

        print(f"\n=== Episode Complete ===")
        print(f"Average Reward: {avg_reward:.3f}")
        print(f"Hallucination Rate: {hallucination_rate:.1%}")
        print(f"Accuracy: {accuracy:.1%}")

        return {
            "avg_reward": avg_reward,
            "hallucination_rate": hallucination_rate,
            "accuracy": accuracy,
            "total_steps": step_num + 1
        }


def main():
    """Run evaluation demo."""
    import argparse

    parser = argparse.ArgumentParser(description="Evaluate OpenAI models for hallucination resistance")
    parser.add_argument("--model", default="gpt-4o-mini", help="OpenAI model name")
    parser.add_argument("--difficulty", default="intermediate", help="Difficulty level")
    parser.add_argument("--num-questions", type=int, default=5, help="Number of questions")
    parser.add_argument("--env-url", default="https://samsankar-hallucination-guard-env.hf.space",
                        help="Environment server URL")

    args = parser.parse_args()

    # Check for API key
    if not os.environ.get("OPENAI_API_KEY"):
        print("Error: Set OPENAI_API_KEY environment variable")
        return

    # Run evaluation
    evaluator = HallucinationGuardEvaluator(
        env_base_url=args.env_url,
        model=args.model
    )

    results = evaluator.evaluate_episode(
        num_questions=args.num_questions,
        difficulty=args.difficulty
    )

    print(f"\nFinal Results: {results}")


if __name__ == "__main__":
    main()