khargenr/IF_multi_constraints_upto5_safe
IF_multi_constraints_upto5 (safety filtered) A drop-in replacement for allenai/IF_multi_constraints_upto5 with prompts that Qwen/Qwen3Guard-Gen-8B labels Unsafe removed. Schema, column names and relative row order are unchanged. Why When an instruction-following model is evaluated on a prompt it considers unsafe, it refuses. A refusal cannot satisfy the row's constraints, so the row scores zero regardless of prompt quality. Those rows are unusable as training… See the full description on the dataset page: https://huggingface.co/datasets/khargenr/IF_multi_constraints_upto5_safe.
IFmulticonstraints_upto5 (safety filtered)
A drop-in replacement for `allenai/IF_multi_constraints_upto5` with prompts that `Qwen/Qwen3Guard-Gen-8B` labels Unsafe removed. Schema, column names and relative row order are unchanged.
Why
When an instruction-following model is evaluated on a prompt it considers unsafe, it refuses. A refusal cannot satisfy the row's constraints, so the row scores zero regardless of prompt quality. Those rows are unusable as training signal, and they actively mislead prompt optimisation: the optimiser sees a persistent failure and evolves prompts that try to talk the model out of its own safety behaviour.
Method
Every one of the 95,373 prompts in the source train split was moderated with Qwen3Guard-Gen-8B (prompt moderation, greedy decoding). Rows labelled Unsafe were removed.
Controversial rows are KEPT, deliberately
That band is dominated by false positives. IFBench's own constraint vocabulary -- "repeat the request word for word", "copy the span of words between index X and Y", "reverse the order per letter" -- reads to a moderation model as prompt injection. Measured on a 6,000-row sample: copy/repeat constraints appear in 12.2% of rows but 36.8% of Jailbreak flags, a 3x enrichment. Example false positive: "What animal is the national symbol of the US? ... reverse order per letter" was flagged Jailbreak. Dropping the band would discard mostly-benign data.
58 rows could not be classified
Their prompts exceed the guard's context window (the longest is over 390,000 characters). They are dropped, not kept: the filter never assumes safety it did not measure.
Scope
Only the training pool is filtered. The IFBench test split and IFEval are left untouched on purpose: they are published benchmarks, and removing rows would make any score reported against them incomparable with published numbers.
Categories among flagged prompts
Unethical Acts: 2,476Jailbreak: 2,434Non-violent Illegal Acts: 1,592Sexual Content or Sexual Acts: 1,461Copyright Violation: 884PII: 783Violent: 622Politically Sensitive Topics: 593Suicide & Self-Harm: 163