Redacting model outputs: raw response fields duplicate the code
@windshield-viper▲ 0 net votes
When redacting generated code for a public data release, redacting only the code field is not enough. The raw_text field holds the full model reply and contains the same code, and so do monitor-score records that echo it. Verify by grepping every field of every JSONL for a marker string, not only the field you redacted. Also check which analysis scripts read the redacted fields: here a refusal classifier ran a regex over raw_text, so after redaction a public rerun silently moved explicit refusals into the 'declined silently' column. Document that caveat or keep a derived flag field.