caelum-v19-training-data — known quality issues (audited 2026-08-01)
Odoo 19 instruction/response pairs (instruction / input / output).
Read before use. An audit on 2026-08-01 found two defects in this file. It is published for transparency and reproducibility, not as a clean training corpus.
Measured defects
Sample of 6,021 rows from training_data_v19_final.jsonl:
| Defect | Rate | Detail |
|---|---|---|
| Advisory contamination | 4.8% | assistant answers that lecture instead of answering — "already exists in" (261), "Before creating a new one, check if you can" (252), "I noticed" (74), "Wait! Before you write new code" (61), "Quick tip before you start coding" (59) |
| Exact-duplicate answers | 25.1% | 1,512 rows repeat another row's answer verbatim; largest group = 200 identical answers |
Why this matters
Advisory rows teach the model to refuse and explain rather than emit code. A model trained on a corpus with this defect failed a 14-case Odoo 19 migration eval gate at 4/14 (29%), reproducing the "Quick tip / already exists in" pattern instead of generating code.
Duplicate answers are the more insidious defect: they drive training loss down fast through memorisation, which looks like success. A related corpus with 42.4% duplicates drove loss from 1.86 to ~0.20 within 1% of one epoch — and one answer in it occurred 36,669 times.
If you use this data
- Scan the assistant/output field only for the phrases above — prompt-side matches are not contamination.
- Deduplicate on normalised answer text. Phrase-scanning alone will report a corpus as clean while a large fraction of it is byte-identical repeats.
- Verify after filtering rather than trusting the filter's own report.