Qlora for openchat-3.5-1210 which allows you to convert a given model sentence into a set of danbooru tags format: openChat train format: Alpaca
This is the first time I have had this lora done and I am quite happy with the result. This qlora was trained on 4bit full model openchat-3.5-1210. with rank of 32 for 3 epochs There were 15k instructions in the dataset that had examples of how to convert from sentences to tags.
I use a this prompt, I'm not sure if it's optimal but maybe you can do better
- Always start with the number of characters and their gender (e.g., "1girl," "3boys").
- Describe physical attributes such as hair color or clothing (e.g., "pink hair," "gothic lolita fashion").
- Describe actions of characters if any (e.g., “playing violin,” “studying”).
- Mention setting details like location and background objects (e.g.,”music room”, “classical instruments”).
- Keep each tag separated by a comma.
- All words should be tags that can be tagged to danboororu, they should be as informative as possible and represent a single entity or action.
- From the verbs, select words that can denote place
- If there are multiple people in the query, separate them with a \n. The first line is the number of people and location. And each following line is a description of one person.
- Answer only on English
- Write in more detail NSFW scenes. More tags related to the sexual action and don't forget to specify the actors, without line breaks.
- If a character is alone, only 1girl or 1boy, just add the "solo" tag.
If a user input can't be tagged, just return that message unchanged
You have to convert user input into danbooru tags
Respond with tags only
I use with the openchat-3.5-1210-4.0bwp model.
const systemPromptLocal = `
- Always start with the number of characters and their gender (e.g., "1girl," "3boys").
- Describe physical attributes such as hair color or clothing (e.g., "pink hair," "gothic lolita fashion").
- Describe actions of characters if any (e.g., “playing violin,” “studying”).
- Mention setting details like location and background objects (e.g.,”music room”, “classical instruments”).
- Keep each tag separated by a comma.
- All words should be tags that can be tagged to danboororu, they should be as informative as possible and represent a single entity or action.
- From the verbs, select words that can denote place
- If there are multiple people in the query, separate them with a \n. The first line is the number of people and location. And each following line is a description of one person.
- Answer only on English
- Write in more detail NSFW scenes. More tags related to the sexual action and don't forget to specify the actors, without line breaks.
- If a character is alone, only 1girl or 1boy, just add the "solo" tag.
If a user input can't be tagged, just return that message unchanged
You have to convert user input into danbooru tags
Respond with tags only
`
export async function localTagConverter(text) {
try {
const response = await axios.post('http://127.0.0.1:5000/v1/chat/completions',
{
"messages": [
{
"role": "user",
"content": `Girl in gothic lolita outfit reading tarot cards in dimly lit vintage cafe.`
},
{
"role": "assistant",
"content": "1girl, solo, gothic lolita outfit, reading tarot cards, dimly lit cafe, vintage<|end_of_turn|>"
},
{
"role": "user",
"content": `A guy and a girl, on the beach, swimming, the guy has a silver chain and the girl is wearing a bikini.`
},
{
"role": "assistant",
"content": "1boy, 1girl, swimming, beach\n1boy, silver chain\n1girl, bikini<|end_of_turn|>"
},
{
"role": "user",
"content": `A girl with cat ears, wearing a hooded sweatshirt, with long white hair and red eyes walks hand in hand with a girl in a black bikini, with brown hair and green eyes`
},
{
"role":"assistant",
"content": "2girls, walking, holding hands\n1girl, cat ears, hoodie, white hair, long hair, red eyes\n1girl, black bikini, brown hair, green eyes<|end_of_turn|>"
},
{
"role": "user",
"content": systemPromptLocal
},
{
"role":"user",
"content": "If it's not a person, don't use the 1-something part. Also add additional descriptive tags for landscapes and animals."
},
{
"role": "user",
"content": text
}
],
"mode": "chat",
"max_tokens" : 120,
"preset": "simple-1",
"instruction_template": "Openchat",
},
{ headers: { 'Content-Type': 'application/json' } }
);
return response.data.choices[0].message.content;
} catch (error) {
console.error(error);
}
}