Hey again everyone, Following up on my earlier posts about converting a visual novel script into a fine-tuning dataset, I\u2019ve gone back and improved the format significantly thanks to feedback here.<\/p>\n
The goal is the same: create expressive, roleplay-friendly dialogue data that captures emotion, tone, character personality, and nuance, especially for dere-type characters and NSFW\/SFW variation.<\/p>\n
VOl 0 is only SFW <\/p>\n
\u2022 What\u2019s New:<\/p>\n
Improved JSON structure, closer to ShareGPT format<\/p>\n
More consistent tone\/emotion tagging<\/p>\n
Added deeper context awareness (4 lines before\/after)<\/p>\n
Preserved expressive elements (onomatopoeia, stutters, laughs)<\/p>\n
Categorized dere-type and added voice\/personality cues<\/p>\n
\u2022 Why?<\/p>\n
Because tagging a line as just \u201claughing\u201d misses everything. Was it sarcasm? Pain? Joy? I want models to understand motivation and emotional flow \u2014 not just parrot words.<\/p>\n
Example (same as before to show improvement):<\/p>\n
Flat version:<\/p>\n
{ “instruction”: “What does Maple say?”,<\/p>\n
“output”: “Oopsie! I accidentally splashed some hot water on you! Sorry about that~ Ahahah– Owwww!!”,<\/p>\n
“metadata”: { “character”: “Maple”, “emotion”: “laughing” <\/p>\n
“tone”: “apologetic” }<\/p>\n
}<\/p>\n
\u2022 Updated version with context:<\/p>\n
{ \"from\": \"char_metadata\", \"value\": { \"character_name\": \"Azuki\", \"persona\": \"Azuki is a fiery, tomboyish...\", \"dere_type\": \"tsundere\", \"current_emotion\": \"mocking, amused, pain\", \"tone\": \"taunting, surprised\" } }, { \"from\": \"char\", \"value\": \"You're a NEET catgirl who can only eat, sleep, and play! Huehuehueh, whooaaa!! Aagh, that's hotttt!!!\" }, { \"from\": \"char_metadata\", \"value\": { \"character_name\": \"Maple\", \"persona\": \"Maple is a prideful, sophisticated catgirl...\", \"dere_type\": \"himidere\", \"current_emotion\": \"malicious glee, feigned innocence, pain\", \"tone\": \"sarcastic, surprised\" } }, { \"from\": \"char\", \"value\": \"Oopsie! I accidentally splashed some hot water on you! Sorry about that~ Ahahah-- Owwww!!\" }, { \"from\": \"char_metadata\", \"value\": { \"character_name\": \"Azuki\", \"persona\": \"Azuki is a fiery, tomboyish...\", \"dere_type\": \"tsundere\", \"current_emotion\": \"retaliatory, gleeful\", \"tone\": \"sarcastic\" } }, { \"from\": \"char\", \"value\": \"Heh, my bad! My paw just flew right at'cha! Hahaha!\" } <\/code><\/pre>\n\u2022 Outcome<\/p>\n
This dataset now lets a model:<\/p>\n
Match dere-type voices with appropriate phrasing<\/p>\n
Preserve emotional realism in both SFW and NSFW contexts<\/p>\n
Move beyond basic emotion labels to expressive patterns (tsundere teasing, onomatopoeia, flustered laughter, etc.)<\/p>\n
It\u2019s still a work in progress (currently ~3MB, will grow, dialogs only without JSON yet), and more feedback is welcome. Just wanted to share the next step now that the format is finally usable and consistent.<\/p>\n<\/div>\n
submitted by \/u\/Akowmako <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
