Hallucination Testing Animate

← Back to topic

🎞 ANIMATE · Mode 3 of 5 · +100 XP— Hallucination Testing

📋Judge Model

⚙️Claims

🔄Audit

✅Rate

🎯 Next: Test →

Step 1 / 4

Speed

What's happening

LLM-as-Judge

Use a strong judge model (GPT-4o) to verify factual consistency claim-by-claim

💡 Analogy

Like a fact-checker comparing an article's claims against primary sources.

check_factual_consistency uses GPT-4o as judge. Given a claim and source, the judge returns: supported (claim is in source), contradicted (source says otherwise), or unverifiable (not enough info).

Console Output

> Claim: "Python was created in 1991"
> Source: "Python was first released in 1991"
> Verdict: supported, confidence: 0.99
> Reasoning: "Source explicitly states 1991"

●hallucination_test.py

LLM-as-JudgeL5–15

1from openai import OpenAI

2import json

4client = OpenAI()

6# Factual consistency check using LLM-as-judge

7def check_factual_consistency(claim: str, source: str) -> dict:

8 prompt = f"""Check if the claim is factually supported by the source.

9Return JSON: {{"verdict": "supported|contradicted|unverifiable",

10 "confidence": 0.0-1.0, "reasoning": "..."}}

12Claim: {claim}

13Source: {source}"""

15 response = client.chat.completions.create(

16 model="gpt-4o", # use strong judge model

17 messages=[{"role": "user", "content": prompt}],

18 response_format={"type": "json_object"},

19 )

20 return json.loads(response.choices[0].message.content)

22# Test RAG pipeline for hallucinations

23def audit_rag_response(question: str, context: str, answer: str) -> dict:

24 # Decompose answer into claims

25 claims_prompt = f"List the factual claims in this answer as a JSON array of strings:\n{answer}"

26 claims_response = client.chat.completions.create(

27 model="gpt-4o-mini",

28 messages=[{"role": "user", "content": claims_prompt}],

29 response_format={"type": "json_object"},

30 )

31 claims = json.loads(claims_response.choices[0].message.content)["claims"]

33 results = []

34 for claim in claims:

35 result = check_factual_consistency(claim, context)

36 results.append({"claim": claim, **result})

38 hallucination_rate = sum(1 for r in results if r["verdict"] != "supported") / len(results)

39 return {"claims": results, "hallucination_rate": hallucination_rate}

39 lines

Steps · 4

What's happening

LLM-as-Judge

Use a strong judge model (GPT-4o) to verify factual consistency claim-by-claim

💡 Analogy

Like a fact-checker comparing an article's claims against primary sources.

check_factual_consistency uses GPT-4o as judge. Given a claim and source, the judge returns: supported (claim is in source), contradicted (source says otherwise), or unverifiable (not enough info).

Console Output

> Claim: "Python was created in 1991"
> Source: "Python was first released in 1991"
> Verdict: supported, confidence: 0.99
> Reasoning: "Source explicitly states 1991"

Was this page helpful?

Discussion

← Back to topic

🎞 ANIMATE · Mode 3 of 5 · +100 XP— Hallucination Testing

📋Judge Model

⚙️Claims

🔄Audit

✅Rate

🎯 Next: Test →

Step 1 / 4

Speed

What's happening

LLM-as-Judge

Use a strong judge model (GPT-4o) to verify factual consistency claim-by-claim

💡 Analogy

Like a fact-checker comparing an article's claims against primary sources.

check_factual_consistency uses GPT-4o as judge. Given a claim and source, the judge returns: supported (claim is in source), contradicted (source says otherwise), or unverifiable (not enough info).

Console Output

> Claim: "Python was created in 1991"
> Source: "Python was first released in 1991"
> Verdict: supported, confidence: 0.99
> Reasoning: "Source explicitly states 1991"

●hallucination_test.py

LLM-as-JudgeL5–15

1from openai import OpenAI

2import json

4client = OpenAI()

6# Factual consistency check using LLM-as-judge

7def check_factual_consistency(claim: str, source: str) -> dict:

8 prompt = f"""Check if the claim is factually supported by the source.

9Return JSON: {{"verdict": "supported|contradicted|unverifiable",

10 "confidence": 0.0-1.0, "reasoning": "..."}}

12Claim: {claim}

13Source: {source}"""

15 response = client.chat.completions.create(

16 model="gpt-4o", # use strong judge model

17 messages=[{"role": "user", "content": prompt}],

18 response_format={"type": "json_object"},

19 )

20 return json.loads(response.choices[0].message.content)

22# Test RAG pipeline for hallucinations

23def audit_rag_response(question: str, context: str, answer: str) -> dict:

24 # Decompose answer into claims

25 claims_prompt = f"List the factual claims in this answer as a JSON array of strings:\n{answer}"

26 claims_response = client.chat.completions.create(

27 model="gpt-4o-mini",

28 messages=[{"role": "user", "content": claims_prompt}],

29 response_format={"type": "json_object"},

30 )

31 claims = json.loads(claims_response.choices[0].message.content)["claims"]

33 results = []

34 for claim in claims:

35 result = check_factual_consistency(claim, context)

36 results.append({"claim": claim, **result})

38 hallucination_rate = sum(1 for r in results if r["verdict"] != "supported") / len(results)

39 return {"claims": results, "hallucination_rate": hallucination_rate}

39 lines

Steps · 4

What's happening

LLM-as-Judge

Use a strong judge model (GPT-4o) to verify factual consistency claim-by-claim

💡 Analogy

Like a fact-checker comparing an article's claims against primary sources.

check_factual_consistency uses GPT-4o as judge. Given a claim and source, the judge returns: supported (claim is in source), contradicted (source says otherwise), or unverifiable (not enough info).

Console Output

> Claim: "Python was created in 1991"
> Source: "Python was first released in 1991"
> Verdict: supported, confidence: 0.99
> Reasoning: "Source explicitly states 1991"