Commit
•
e753120
1
Parent(s):
4655e6d
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
## Captain Adventure 32B
|
| 2 |
+
(Model card will be updated further in the near future.)
|
| 3 |
+
|
| 4 |
+
Fine-tuned from [c4ai-command-r-08-2024](https://huggingface.co/CohereForAI/c4ai-command-r-08-2024) using QLoRA at 8k context for training.
|
| 5 |
+
|
| 6 |
+
Datasets used:
|
| 7 |
+
- Spring Dragon
|
| 8 |
+
- Skein
|
| 9 |
+
|
| 10 |
+
Total was around ~29-30M tokens of classic-style text adventure data.
|
| 11 |
+
|
| 12 |
+
This was trained in `completion` mode on axolotl, meaning I did not introduce any instruct format.
|
| 13 |
+
|
| 14 |
+
The intended use is in the style of KoboldLite's Adventure mode, which is raw text completion of the form:
|
| 15 |
+
```
|
| 16 |
+
Story text, setup, words
|
| 17 |
+
|
| 18 |
+
> User action
|
| 19 |
+
|
| 20 |
+
Narration describing the outcome of this action
|
| 21 |
+
|
| 22 |
+
> Another user action
|
| 23 |
+
|
| 24 |
+
More narration from the bot...
|
| 25 |
+
```
|
| 26 |
+
|
| 27 |
+
Set `>` as a stop token to use in traditional completion format.
|
| 28 |
+
|
| 29 |
+
**I've also been told this works well in instruct mode, in which case you will want to use Command-R Instruct format.**
|
| 30 |
+
|
| 31 |
+
Thanks, hope you have some fun with it.
|
| 32 |
+
|
| 33 |
+
*Extra thanks to Gargamel for sponsoring compute for this train. If you'd like me to train more models like this, feel free to donate to [my ko-fi](https://ko-fi.com/toastypigeon) so I can buy more compute credits.*
|