FT-Fail: saved fine-tuning diagnosis cases

FT-Fail contains evidence from the Next Experiment project. This release packages completed measurements. It contains 13 historical Token Factory cases, their 52 method score rows, and a separate 200-query automatic label audit with 400 reference queries.

Configurations

Configuration	Rows	Content
cases	5 dev + 8 historical held-out	Symptoms, recipe, checkpoint losses, planted faults and saved quality
methods	20 dev + 32 historical held-out	Saved diagnosis, ranking, fix, cost and failure scores for four methods
semantic_audit	200 dev	Query, observed label, source label and recorded corruption
semantic_reference	400 dev	Separate reference queries with original labels

The case input is stored in case_json; the planted truth stays in injected_faults. Evaluate a diagnosis using only the case input. Local artifact paths in the original manifests identify historical files and are not download links. Model checkpoints and all original training artifacts are outside this release.

The historical held-out run finished before publication. Its evidence is already published in the repository and is included here without rerunning methods. These cases are now public and cannot serve as an untouched final evaluation for further development. No pending experiment's final split is included.

Findings and limits

The historical benchmark covers BANKING77, Qwen3-0.6B and predominantly low learning-rate faults. It is too small to establish performance across tasks, models or fault classes. A beneficial repair need not identify the original cause: increasing rank also improved several low-learning-rate cases. The benchmark report describes the fixed decision rule, failed repairs and method comparisons.

The automatic semantic audit uses eight intents from BANKING77's training split. Of its 200 queries, 100 retain their labels, 50 receive a different valid label, and 50 receive an invalid JSON-wrapped label. The original human-review requirement was replaced by an explicitly recorded automatic-reference acceptance amendment. The labels are source labels plus known modifications; they are not independent human judgments.

Nano's explicit contradiction flag found none of the 50 wrong labels. Super found 31, with two HTTP failures retained. Neighbor and cleanlab comparisons, including false positives, are documented in the semantic report. The fixed semantic duplicate threshold missed all four authored paraphrase controls. These measurements do not justify automatic data deletion or calibrated production fault probabilities.

Provenance and reuse

manifest.json records SHA-256 hashes and the source repository revision. The evidence/ directory includes original case manifests, protocols, acceptance amendment and reports. Packaging does not issue new model calls or change scores. The code that constructs this release is scripts/package_ft_fail.py in the source repository.

BANKING77 queries come from PolyAI's task-specific dataset, pinned to commit 57ec275d8078af65b7731c2a98be812d844a6d6b for the automatic audit. See the dataset card and Casanueva et al. (2020). The automatic audit records original source line numbers. Changes include query selection, label replacement and label wrapping; query text is retained.

See LICENSES.md for source attribution and the separate terms for BANKING77 text and project-authored code, manifests and measurements.