Generate an eval dataset from an existing set of chunks.
Positive examples: first 80 chars of chunk content as synthetic query.
Negative examples: queries whose target is far from other chunks (low Jaccard).
Paraphrases: if a judge is provided, rewrites a fraction of queries.
Generate an eval dataset from an existing set of chunks.
Positive examples: first 80 chars of chunk content as synthetic query. Negative examples: queries whose target is far from other chunks (low Jaccard). Paraphrases: if a judge is provided, rewrites a fraction of queries.