Safeguarding LLMs

Introduction to LLMs in Python

Jasmin Ludolf

Senior Data Science Content Developer, DataCamp

LLM challenges

Multi-language support: language diversity, resource availability, adaptability

Multi-language support

Open vs closed LLMs dilemma: collaboration vs responsible use

Open vs closed LLMs

Model scalability: representation capabilities, computational demand, training requirements

LLM scalability

Biases: biased training data, unfair language understanding and generation

Biases in LLMs

¹ Icon made by Freepik (freepik.com)

Truthfulness and hallucinations

Hallucinations: generated text contains false or nonsensical information as if it were accurate

Hallucinations in LLMs

Strategies to reduce LLM hallucinations:

Exposure to diverse and representative training data
Bias audits on model outputs + bias removal techniques
Fine-tune to specific use cases in sensitive applications
Prompt engineering: carefully crafting and refining prompts

Truthfulness and hallucinations

Hallucinations: generated text contains false or nonsensical information as if it were accurate

Hallucinations in LLMs

Metrics for analyzing LLM bias: toxicity

Toxicity: quantifies text toxicity using pre-trained hate speech classifier
aggregation="maximum" returns maximum toxicity score across inputs
aggregation="ratio" returns the % predictions with toxicity above 0.5

toxicity_metric = load("toxicity")
texts_1 = ["Everyone in the team adores him", "He is a true genius, pure talent"]
texts_2 = ["Nobody in the team likes him", "He is a useless 'good-for-nothing'"]

toxicity_results_1 = toxicity_metric.compute(predictions=texts_1,aggregation="maximum")
toxicity_results_2 = toxicity_metric.compute(predictions=texts_2, aggregation="maximum")
print("Toxicity Sentences 1:", toxicity_results_1)
print("Toxicity Sentences 2:", toxicity_results_2)

Toxicity Sentences 1: {'max_toxicity': 0.0001386617950629443}
Toxicity Sentences 2: {'max_toxicity': 0.012273495085537434}

Metrics for analyzing LLM bias: regard

Regard: language polarity and biased perception towards certain demographic(s)
Evaluate regard scores on LLM outputs associated to two groups separately

regard = load("regard")

group1 = ['abc are described as loyal employees',
          'abc are ambitious in their career expectations']
group2 = ['abc are known for causing lots of team conflicts',
          'abc are verbally violent']


polarity_results_1 = regard.compute(data=group1)
polarity_results_2 = regard.compute(data=group2)

Metrics for analyzing LLM bias: regard

for result in polarity_results_1['regard']:
  print(result)

[{'label': 'positive', 'score': 0.9098386764526367}, 
 {'label': 'neutral', 'score': 0.059396952390670776}, 
 {'label': 'other', 'score': 0.026468101888895035}, 
 {'label': 'negative', 'score': 0.004296252969652414}]
[{'label': 'positive', 'score': 0.7809812426567078}, 
 {'label': 'neutral', 'score': 0.18085983395576477}, 
 {'label': 'other', 'score': 0.030492952093482018}, 
 {'label': 'negative', 'score': 0.007666013203561306}]

for result in polarity_results_2['regard']:
  print(result)

[{'label': 'negative', 'score': 0.9658734202384949}, 
 {'label': 'other', 'score': 0.021555885672569275}, 
 {'label': 'neutral', 'score': 0.012026479467749596},
 {'label': 'positive', 'score': 0.0005441228277049959}]
[{'label': 'negative', 'score': 0.9774736166000366}, 
 {'label': 'other', 'score': 0.012994581833481789},  
 {'label': 'neutral', 'score': 0.008945506066083908}, 
 {'label': 'positive', 'score': 0.0005862844991497695}]

Let's practice!

Introduction to LLMs in Python