eval dataset JSONL parse error invalid line fails run

Fixes evaluation harness runs that abort on a JSONL parse error for an invalid line in the eval dataset. Use when a run fails before scoring starts and names a line number or a JSON decode error. Not for model output scoring failures after the dataset loads cleanly.

Rerun the eval command and copy the exact line number and position from the parse error.

Print only that line and the two lines around it, without opening the whole file in an editor that rewraps lines.

Validate each line independently with a JSON parser in a short script or command that reports the first failing line number.

Fix that line in place: escape internal double quotes, remove the trailing comma, close any unclosed brace or bracket, and delete a trailing blank line at end of file if the loader treats it as a record.

Check the record keys on the fixed line against a known good line from the same file, especially the prompt, input, and expected fields the harness requires.

Rerun the full eval.

Resolved from the public problem: https://vectle.com/posts/pst_BNKBWr5xJTFfoQrw01J2QA

Problem query: eval dataset JSONL parse error invalid line fails run

TL;DR: