Skip to content

Checks

dropoutt 1.4.1 includes 36 checks. Identifiers are never renumbered. Run dropoutt checks to list them or pass a check id for a detailed description.

Tier 0: structural

Tier 0 checks identify exact structural and formatting defects.

IDDescription
T0-SCHEMA-001Files are not training data: agent session logs and telemetry that look like chat
T0-SCHEMA-002One folder mixes several record layouts
T0-SCHEMA-003Records failed to parse
T0-SCHEMA-004Message content was not a string
T0-SCHEMA-005Content sits in keys the layout never reads
T0-FORMAT-001Plain-text files are holding structured records
T0-GEN-001Generator scaffolding outside the records
T0-REASON-001Only some responses carry a reasoning trace
T0-TRUNC-002Responses stop at a generation length cap
T0-QUAL-001Documents whose lines mostly lack punctuation
T0-QUAL-002Documents built mostly from very short lines
T0-QUAL-003Documents repeating their own lines
T0-ROLE-001Conversation role structure is invalid
T0-ROLE-002Role names are not the canonical vocabulary
T0-TMPL-001Data is already formatted with a chat template
T0-TMPL-002Records fail to render with the target chat template
T0-MASK-001Records contribute zero trainable tokens
T0-MASK-002Stop token is outside the trainable span
T0-TRUNC-001Records exceed the sequence length
T0-PACK-001Packing efficiency under concat-and-chunk
T0-ENC-001Text encoding is damaged: mojibake from a file decoded as the wrong codec
T0-DUP-001Exact and whitespace-identical duplicates
T0-DEGEN-001Degenerate responses

Tier 1: statistical

IDDescription
T1-NDUP-001Near-duplicate records (MinHash; reports, never deletes)
T1-DUP-002Same prompt answered two different ways
T1-OVERLAP-001Datasets overlap with each other, directionally
T1-ATLAS-001Corpus sits in very few topical regions, dropoutt atlas only
T1-ATLAS-002A crowded region holds near-identical records, dropoutt atlas only
T1-CONTAM-001Training data overlaps evaluation benchmarks
T1-LANG-001Language composition and detection confidence
T1-LANG-002Records deviate from the dataset’s main language
T1-LANG-003Script does not match the detected language: Turkish that lost its diacritics reads as a different language
T1-PII-001Personal data and credentials in training text
T1-IDENT-001Assistant identity leakage and refusal boilerplate
T1-STYLE-001Formulaic response openings
T1-LIC-001Datasets have no recorded licence

Behavior and scope

  • Blocking requires an explicit --target.
  • Checks report recommended actions but do not delete or rewrite data.
  • PII matches are masked in reports.
  • PDF text extraction is not supported. Scan extracted text instead.
  • Nothing phones home. Network access is limited to fetching tokenizers and the atlas encoder, and --offline disables it.
Every finding is labelled unverified: no calibration corpus yet links acting on a finding to a measured change in model quality. Structural checks identify deterministic defects such as empty loss masks and fully truncated responses; treat statistical findings as diagnostic signals and validate changes against the target model and evaluation suite.