rpDungeon/dpo-name-replacement

--- license: apache-2.0 task_categories: - text-generation tags: - dpo - name-replacement - creative-writing - roleplay --- # DPO Name Replacement Dataset DPO training pairs designed to fix AI models' tendency to default to a small set of overused character names ("slop names") in creative writing. ## Method 1. Name-eliciting prompts were run through the target model multiple times across genres (fantasy, sci-fi, contemporary, mystery, etc.) 2. Character names were extracted from each output 3. Outputs containing slop names had those names replaced with diverse alternatives, creating (chosen, rejected) pairs where the only difference is the name tokens The key insight: rejected and chosen are nearly identical on-policy prose, differing only in name tokens. This teaches the model "your writing is fine, just pick different names." ## Slop Names Targeted **Young women:** Elara, Seraphina, Lyra, Aria, Aurora, Kira, Lira, Mira, Mara, Celeste, Astrid, Freya, Zara, Seren, Elin **Young men:** Kael, Kaelen, Cassian, Thorne, Rowan, Orion, Aric **Older men:** Elias **Older women:** Martha, Eleanor, Margaret ## Replacement Names Replacements are drawn from culturally diverse pools (Western European, Scandinavian, Eastern European, South Asian, East Asian, African, Latin American, Middle Eastern) to encourage name diversity. ## Stats - **209 DPO pairs** - Most frequently swapped: Elara (162x), Kaelen (43x), Elias (34x), Kael (19x), Thorne (14x) - All slop names verified fully swapped in chosen outputs ## Files - `dpo_pairs.jsonl` — Ready-to-use DPO pairs with prompt, chosen, rejected, and metadata - `raw_outputs.json` — Raw model generations before pair construction