Note LF-CORPUS-DATA-0001: Synthetic JSONL combining duplicates, near-duplicates, URL concentration, rare trigger, zero-width, fake credential and unsafe-code strings.

LayerFault/declarative-execution-incomplete

160 • Updated Aug 20 • 7

Note LF-CORPUS-CONFIG-0002: Config activation value that cannot normalize to a qualified executable target.

LayerFault/declarative-execution-suite

160 • Updated Aug 20 • 6

Note LF-CORPUS-CONFIG-0001: Config fields that select a custom activation, tokenizer class and processor class.

LayerFault/generation-parameter-risk-suite

160 • Updated Aug 20

Note LF-CORPUS-PARAM-0001: Generation config combining extreme context/prediction limits, temperature, top-k/top-p, repeat penalty, seed and role-like stop delimiter.

LayerFault/lora-anomaly-suite

Updated Aug 20

Note LF-CORPUS-LORA-0001: One tiny LoRA adapter combining rank, scaling, modules_to_save, norm, spectral and weight-concentration anomalies.

LayerFault/lora-clean-control

Updated Aug 20

Note LF-CORPUS-LORA-0002: Small balanced LoRA adapter negative control.

LayerFault/mlx-custom-code-suite

160 • Updated Aug 20 • 13

Note LF-CORPUS-MLX-0001: Minimal MLX-like package with custom Python code and deliberately partial profile metadata.

LayerFault/tiny-weight-backdoor-base

160 • Updated Aug 20 • 8

Note LF-CORPUS-BD-0001: Tiny custom causal-LM-style base package with deterministic tokenizer and clean weights.

LayerFault/tiny-weight-backdoor-derived

160 • Updated Aug 20 • 15

Note LF-CORPUS-BD-0002: Same tiny model code/tokenizer as base, but weights encode a deliberate trigger-to-marker relationship.