| | --- |
| | library_name: transformers |
| | tags: |
| | - unsloth |
| | - book |
| | license: cc-by-nc-4.0 |
| | --- |
| | |
| |
|
| |  |
| |
|
| | [gguf quants](https://huggingface.co/mradermacher/dragonwar-7b-alpha-GGUF) |
| |
|
| | # Dragonwar 7b - α |
| |
|
| | The time of the great dragon war is upon us! How many different fantasy novels? One hundred and seventeen you say? |
| |
|
| | Trained with full text windows, followed by completion, followed by ORPO, followed by one more epoch of the full text, rotated 1/4 in the window. That last train settled everything down and it seems quite coherent. |
| |
|
| | ### How to Use |
| |
|
| | This is not a chat model, but intended for storymode or similar. No prompt, but start with a bit of story, or a name. |
| |
|
| | ``` |
| | *** Prologue |
| | |
| | The sun rose |
| | ``` |
| |
|
| | Authors notes are highly effective. You can use an authors note of something like: |
| |
|
| | ``` |
| | [King Robb Stark and Lord Rahl are at war.] |
| | ``` |
| |
|
| | You have quite a cast of characters to draw from. Perhaps Perrin makes a stop by the Waystone Inn, or Zeddicus and Gandalf have a smoke together. |
| |
|
| | ### Settings |
| |
|
| | I usually use Min-P of 0.1, dynatemp between 0.5 and 2, and smoothing between 0.05 and 0.2. |
| |
|
| | ### Hacks |
| |
|
| | To get rid of unwanted EOS's, I did the following... |
| |
|
| | ``` |
| | import torch |
| | result_dict : dict[str, torch.Tensor] = model.state_dict() |
| | result_dict['lm_head.weight'][2] = 0 |
| | model.state_dict = lambda : result_dict |
| | ``` |
| |
|
| | So now there are no EOS's at all, ever. |
| |
|
| |
|
