How should I practice Excel Training Data and Model Response Audit?

Start by reading the full statement, solve the visible examples, then compare the pattern against related Microsoft phone screen problems in FastPrep.

Is Excel Training Data and Model Response Audit associated with Microsoft?

FastPrep tags this as a Microsoft phone screen coding practice problem when the problem catalog has that company and stage context.

Excel Training Data and Model Response Audit

Microsoft● EasyFULLTIMEPHONE SCREEN

Learn

Problem statement

Audit training records for an Excel assistant, then evaluate the recorded responses of a pretrained model on the valid records. Return the record counts and a decision under the threshold policy below.

Records and normalization

The arrays prompts, tables, expectedResponses, and modelResponses have the same length N. Index i describes one record.
tables[i] is that record's spreadsheet, represented as rows of integer cells. A table may be empty or ragged; these are conditions the audit must detect.
expectedResponses[i] is the reference answer, and modelResponses[i] is an already recorded model prediction. Do not run a model.
Trim only leading and trailing space (' '), tab ('\t'), carriage return ('\r'), and newline ('\n'). Preserve case and every interior character.
Count every input record separately, including duplicates.

Valid records and correct predictions

A training record is valid if and only if all three conditions hold:

Its trimmed prompt is nonempty.
Its trimmed reference answer is nonempty.
Its table has at least one row, every row has the same number of cells, and that common number is positive.

For valid records, treat the supplied reference answers as correct. The audit checks the table's structure rather than recomputing the reference answer from its cells. A model prediction is correct exactly when its trimmed text equals the trimmed reference, with case and interior whitespace preserved. An empty model prediction is incorrect; it does not make the training record invalid. Exclude invalid records from model evaluation.

Threshold decision

Return [validRecordCount, invalidRecordCount, correctPredictionCount, fineTuneDecision] in that order:

invalidRecordCount = N - validRecordCount.
If there are no valid records, set fineTuneDecision = -1 for insufficient evaluation data, regardless of the threshold.
Otherwise set fineTuneDecision = 1 if 100 * correctPredictionCount < minimumAccuracyPercent * validRecordCount, meaning accuracy is below the required threshold.
Set fineTuneDecision = 0 when the threshold is met. Equality meets the threshold.

This decision is the exercise's fixed accuracy policy for a fine-tuning recommendation.

Function

auditExcelTrainingData(prompts: String[], tables: int[][][], expectedResponses: String[], modelResponses: String[], minimumAccuracyPercent: int) → int[]

Examples

Example 1

prompts = ["sum"," ","first"]tables = [[[1,2]],[[3]],[[5],[6]]]expectedResponses = ["3","3","5"]modelResponses = ["3","3","6"]minimumAccuracyPercent = 60return = [2,1,1,1]

The blank prompt makes the middle record invalid. Among the two valid records, only the first prediction matches, so accuracy is 1/2, below 60%.

Example 2

prompts = ["p","q"]tables = [[],[[1],[2,3]]]expectedResponses = ["a","b"]modelResponses = ["a","b"]minimumAccuracyPercent = 0return = [0,2,0,-1]

The first table has no rows and the second is ragged. There are no valid records to evaluate, even though the threshold is 0%.

Example 3

prompts = ["\t p\n","q"]tables = [[[0]],[[5]]]expectedResponses = ["  Yes\r\n","x"]modelResponses = ["Yes"," x\t"]minimumAccuracyPercent = 100return = [2,0,2,0]

After trimming only the specified outer characters, both predictions exactly match their references. Accuracy equals 100%, so the threshold is met.

Constraints

0 <= N <= 100; all four record arrays have length N.
Each table has 0..10 rows, each row has 0..10 cells, and the dataset contains at most 1000 numeric cells. Ragged rows are permitted as input.
Every numeric cell is an integer in [-100, 100].
Each string has length 0..100. Across all three string arrays there are at most 3000 characters.
Strings contain only printable ASCII characters with codes 32..126, plus tab, carriage return, and newline.
0 <= minimumAccuracyPercent <= 100.