BookAdventures is a research model to serve as a potential all round model with better long form writing, due to our lack of suitable Chat RP data the past months we have chosen to release the intermediate model as is for others to expand upon as other communities have access to superior private chat data. This model was tuned on top of KoboldAI/LLaMA-3.1-8B-Infinity3M-Kobo restoring its writing capability but replacing this writing with a much longer form. In our testing 8B is not ideal for this, but it is the largest we could tune at 32K context.

This model intentionally writes like a book, expect entirely random openings where what you asked for is weaved in to the story. Its designed for guided co-writing with an instruct prompt describing the entire plot summary. We also added our usual Adventures dataset making this double as an adventure mode model, but due to lack of a suitable chat dataset this model is incapable of engaging Chat RP leaving it one step short of our original goal for an all round model.

For the best results use this model in KoboldAI Lite.

This model was trained on the Alpaca format with a large subset of the Infinity3M dataset, it should respond well to alpaca.

Format your prompt as follows:

### Instruction:
Write a GENRE novel about SUMMARY OF THE STORY
### Response:

Including the genre in this position is very important, this is what all our long form example prompts used. You want this instruction in our context menu so it remains visible to the AI. You should now be able to co-write your story with the summary guiding it. Note, the data expects longer summaries that are a good paragraph in size, only giving it a topic will work less well.

Text adventures can be done in two ways, this model has support for our traditional adventure mode and will then behave like the classic versions of AI Dungeon but with many stories to draw inspiration from similar to our old Skein and Nerys models. You can also use instruct mode instucting it to "Start a text adventure about" in which case it will have longer form writing for your adventure.

You will be dissapointed since this model has no such data, look if anyone finetuned such a model on top of this one or succesfully merged this model.

This model to our knowledge used a unique approach to give it a longform writing bias, if you did the same method before us please let us know so we can give you credit. We first stripped Infinity3M of all its short form writing data to prevent the model from ending stories early and to reduce the "slop" that writers often complain about. Then we used our own PromptGen tool to generate instruct prompts for the Pike dataset (Thanks Mr.Seeker for letting me use it for this experiment, it saved a lot of time cleaning book data - Henk) The generated prompts were checked and cleaned, prompts that accidentally featured references to the original works or artists were rewritten or removed to ensure the model could not learn to copy anyones work or style. In generating the data we had roughly a 10% failure rate where LLama3.1-8B-Instruct would not follow the tasks correctly. Many of these could be saved, but we also had to remove a decent amount of stories due to the prompt not generating correctly. Specialized models would help here.

Lastly we added the Floyd adventure data from the Skein model with a light enough bias not to infect the books.

This experiment was only partially succesfull, there is a chance the model looses track before anchoring itself down to your prompt by introducing the story elements in time. To test the model correctly it must be generating longer stories since short stories are not its intended purpose and within the usual 512 tokens other models generate it will almost certainly not have included your story element. Short stories were omitted but could likely be introduced succesfully had it been distinct enough in the data / prompt language. Model has no knowledge of Chat RP. Model will hallucinate incorrect story authors from the base model, in our testing we could trace these back to the gutenberg data present in Llama-3.1. If your name is mentioned this does not mean your work is in our data.

This model follows the Llama-3.1 license / CC-BY-NC-SA-4.0 and is intended as a research only model. We don't mind private use by AI hobbyists, but do not use this model for commercial purposes.