Question answering in Prompt Engineering / GenAI - Model Pipeline Trace

Model Pipeline - Question answering

This pipeline takes a question and a related text passage as input. It processes the text to understand the context, then uses a trained model to find the answer to the question within the passage.

Data Flow - 6 Stages

1Input data

1000 samples x 2 texts (question and passage)→Receive pairs of questions and passages→1000 samples x 2 texts

Question: 'What color is the sky?' Passage: 'The sky is blue on a clear day.'

↓

2Preprocessing

1000 samples x 2 texts→Tokenize and convert texts to numerical tokens→1000 samples x 2 sequences of tokens (max length 128)

Question tokens: [101, 2054, 2079, 2003, 1996, 3710, 102] Passage tokens: [101, 1996, 3710, 2003, 2631, 2006, 1037, 2706, 2154, 1012, 102]

↓

3Feature Engineering

1000 samples x 2 sequences of tokens→Create input embeddings combining question and passage tokens→1000 samples x 128 tokens x 768 features

Embedding vector for token 101: [0.12, -0.05, ..., 0.33]

↓

4Model Training

1000 samples x 128 tokens x 768 features→Train transformer-based QA model to predict answer span→1000 samples x 2 indices (start and end positions)

Predicted answer span: start=4, end=4 (tokens corresponding to 'blue')

↓

5Evaluation

1000 samples x 2 indices→Calculate exact match and F1 score comparing predicted and true answers→Metrics: exact match=0.78, F1=0.85

Exact match: 78% of predicted answers exactly match true answers

↓

6Prediction

1 sample x 2 texts→Model predicts answer span for new question and passage→1 sample x answer text

Question: 'What color is the sky?' Answer: 'blue'

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.9 |***
0.7 |**
0.5 |*
0.4 |

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning, loss high, accuracy low
2	0.9	0.60	Loss decreases, accuracy improves
3	0.7	0.70	Model learns better answer spans
4	0.5	0.78	Loss continues to decrease, accuracy rises
5	0.4	0.82	Model converges with good performance

Prediction Trace - 5 Layers

Layer 1: Tokenization

Layer 2: Embedding Layer

Layer 3: Transformer Encoder

Layer 4: Answer Span Prediction

Layer 5: Answer Extraction

Model Quiz - 3 Questions

What does the model predict as output after training?

A A yes or no answer

B A summary of the passage

C Start and end positions of the answer in the passage

D The question rewritten

Key Insight

This visualization shows how a question answering model learns to find exact answer spans in text passages by converting words into meaningful vectors and improving predictions over training epochs.