LLM Training Dataset Quality Check for Excel Copilot Tasks

Question: LLM Training Dataset Quality Check for Excel Copilot Tasks You are given a dataset dataset used to train/evaluate an LLM. Each sample contains: prompt: the user instruction (string) table: an Excel-like table (2D array/row-column structure; cells are strings or numbers; may include headers) response: the expected output (string) The dataset supports Excel Copilot scenarios such as answering questions from tables, computing aggregations, filtering/sorting, generating formulas, and explaining tables. Design and implement a dataset quality checking program that: Validates structure and parseability prompt/table/response existence, correct types, non-empty. table is rectangular (all rows have the same number of columns). Table size does not exceed limits (rows/cols/total cells). Detects content inconsistencies / anomalies (implement at least 3 categories) Choose at least 3 checks and report the reason when triggered: Duplicate / near-duplicate samples (same prompt+table) response too short/too long or obviously irrelevant Language mismatch between prompt and response, or response does not address the prompt (simple heuristics allowed) Abnormal numeric formatting in the table (many "N/A", empty cells, non-parsable numbers) Potential leakage (e.g., response directly copies a full column/table text when the prompt doesn’t require it) Generates a summary report Overall stats: total, passed/failed, counts per failure reason. A few example bad cases (e.g., first K indices per reason). Assess whether a HuggingFace pretrained model needs fine-tuning (high-level only) Without writing training code, describe how you would use this dataset to evaluate whether a pretrained model can handle the Excel tasks or needs fine-tuning. Include: split strategy, at least 2 evaluation metrics, and how to avoid data leakage. I/O Input: read JSON from stdin: { "dataset": [ {"prompt": "...", "table": [["..."], ["..."]], "response": "..."} ], "k": 3, "max_rows": 200, "max_cols": 50, "max_cells": 5000, "max_response_len": 2000 } Output: print a JSON report, at minimum: { "total": 0, "passed": 0, "failed": 0, "reason_counts": {"...": 0}, "examples": {"reason": [0,1,2]} } Constraints Number of samples N: 1 to 50,000 Max cells per table limited by max_cells Notes Heuristics are acceptable for consistency/anomaly checks. Focus on production usability: extensible checks, explainable reasons, and controlled time complexity. Example Input {"dataset":[{"prompt":"Sum sales in 2024","table":[["Year","Sales"],[2023,10],[2024,20]],"response":"20"}],"k":2,"max_rows":200,"max_cols":50,"max_cells":5000,"max_response_len":2000} Output {"total":1,"passed":1,"failed":0,"reason_counts":{},"examples":{}}