Instruction formatting in Prompt Engineering / GenAI - Model Metrics & Evaluation

Metrics & Evaluation - Instruction formatting

Which metric matters for Instruction formatting and WHY

When working with instruction formatting in AI models, the key metric to focus on is accuracy. This measures how well the model follows the instructions given. If the model misunderstands or misformats the instructions, the output will be wrong or confusing.

Accuracy matters because the goal is to get the model to produce exactly what the instruction asks for. Other metrics like precision or recall are less relevant here because we want the entire instruction to be correctly followed, not just parts of it.

Confusion matrix or equivalent visualization

Instruction Followed Correctly | Instruction Followed Incorrectly
------------------------------|------------------------------
True Positive (TP): Correctly formatted instructions  | False Negative (FN): Instructions not followed
False Positive (FP): Incorrect formatting accepted as correct | True Negative (TN): Not applicable here

Total instructions = TP + FP + FN
    

In instruction formatting, TP means the model output matches the instruction perfectly. FN means the model failed to follow the instruction. FP and TN are less common but can represent cases where incorrect formatting is mistakenly accepted.

Precision vs Recall tradeoff with examples

In instruction formatting, precision means how often the model's formatted output is actually correct when it claims to be correct.

Recall means how many of the instructions the model correctly formats out of all instructions given.

For example, if a model formats 90 outputs and 80 are correct (precision ~89%), but it only correctly formats 80 out of 100 instructions (recall 80%), it means it is careful but misses some instructions.

Depending on the use case, you might want higher recall (follow all instructions even if some are imperfect) or higher precision (only produce output when very sure it is correct).

What "good" vs "bad" metric values look like for instruction formatting

• Good: Accuracy above 95%, precision and recall both high (above 90%). This means the model follows instructions well and rarely makes mistakes.
• Bad: Accuracy below 70%, precision or recall very low (below 50%). This means the model often misunderstands or misformats instructions.
• Balanced precision and recall are important. High precision but low recall means many instructions are ignored. High recall but low precision means many outputs are wrong.

Common pitfalls in instruction formatting metrics

• Accuracy paradox: If instructions are very simple or repetitive, a model might get high accuracy by guessing common patterns but fail on new instructions.
• Data leakage: If the model sees test instructions during training, metrics will be unrealistically high.
• Overfitting: The model might memorize specific instructions but fail to generalize to new ones, causing poor real-world performance.
• Ignoring partial correctness: Sometimes outputs partially follow instructions. Metrics that only count perfect matches miss this nuance.

Self-check question

Your model has 98% accuracy but only 12% recall on following instructions. Is it good for production? Why or why not?

Answer: No, it is not good. The high accuracy likely means the model is correct when it tries, but the very low recall means it only follows a small fraction of instructions. This means many instructions are ignored, which is a big problem for instruction formatting.

Key Result

Accuracy is key for instruction formatting; balanced precision and recall ensure instructions are followed correctly and consistently.