glm-wikisummary-if-it3

# glm-wikisummary-if-it3 25k Wikipedia summary SFT dataset with stacked verifiable constraints augmented onto prompts. ## Changes from it2 - 5k -> 25k base rows - 0-6 stacked constraints per prompt - Constraints are chosen because the completion already satisfies them - ground truth by construction ## Constraints | ID | Instruction | Mutates completion? | |---|---|---| | `no_commas` | "Do not use any commas in your response." | No | | `exact_commas` | "Use exactly N commas in your response." | No | | `no_digits` | "Do not use any numbers or digits in your response." | No | | `no_parens` | "Do not use any parentheses in your response." | No | | `starts_with` | "Begin your response with \`{prefix}\`." | No | | `ends_with` | "End your response with \`{suffix}\`." | No | | `all_upper` | "Write your entire response in uppercase letters." | Yes (`.upper()`) | | `all_lower` | "Write your entire response in lowercase letters." | Yes (`.lower()`) | Incompatible pairs: `{no_commas, exact_commas}`, `{all_upper, all_lower}`. ~17% unconstrained pass-throughs. Rest: 1-6 constraints per sample. ## Previous iterations - [it1](https://huggingface.co/datasets/kalomaze/glm-wikisummary-if-it1): 5k rows, 26 format instructions (word limits, bullet points, etc.) - [it2](https://huggingface.co/datasets/kalomaze/glm-wikisummary-if-it2): same as it1 with fixed template wording