Factual consistency checking in Prompt Engineering / GenAI - Model Pipeline Trace

Model Pipeline - Factual consistency checking

This pipeline checks if the facts in a generated text match the original source. It helps make sure AI outputs are truthful and reliable.

Data Flow - 5 Stages

1Input data

1000 text pairs (source and generated) x 2 columns→Load pairs of original source text and AI-generated text→1000 text pairs x 2 columns

Source: 'The Eiffel Tower is in Paris.' Generated: 'The Eiffel Tower is located in Paris, France.'

↓

2Text preprocessing

1000 text pairs x 2 columns→Clean text, tokenize sentences, remove stop words→1000 text pairs x 2 columns (cleaned tokens)

Source tokens: ['Eiffel', 'Tower', 'Paris'] Generated tokens: ['Eiffel', 'Tower', 'Paris', 'France']

↓

3Feature extraction

1000 text pairs x 2 columns (tokens)→Convert tokens to embeddings using language model→1000 pairs x 2 embeddings (each 768 dimensions)

Source embedding: [0.12, -0.05, ..., 0.33] Generated embedding: [0.10, -0.04, ..., 0.30]

↓

4Model training

1000 pairs x 2 embeddings→Train classifier to predict if generated text is factually consistent→Trained model

Model learns to output 1 if consistent, 0 if not

↓

5Evaluation

200 test pairs x 2 embeddings→Predict consistency and compare with true labels→200 predictions with accuracy metric

Accuracy: 85% correct factual consistency predictions

Training Trace - Epoch by Epoch

Loss
0.7 |    
0.6 |*   
0.5 |**  
0.4 |*** 
0.3 |****
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Model starts learning basic patterns
2	0.50	0.72	Accuracy improves as model learns fact matching
3	0.40	0.80	Model better distinguishes consistent vs inconsistent
4	0.35	0.83	Loss decreases steadily, accuracy improves
5	0.30	0.85	Model converges with good factual consistency detection

Prediction Trace - 4 Layers

Layer 1: Input pair

Layer 2: Text preprocessing

Layer 3: Feature extraction

Layer 4: Model prediction

Model Quiz - 3 Questions

What does the model predict when the generated text matches the source facts?

A Prediction 0 (inconsistent)

B Prediction 2 (unknown)

C Prediction 1 (consistent)

D No prediction

Key Insight

Factual consistency checking uses a model to compare original and generated texts. The model learns to spot if the generated text keeps the facts right by turning words into numbers and training on many examples. As training goes on, the model gets better and more accurate at this task.