Update README.md
Browse files
README.md
CHANGED
|
@@ -11,6 +11,16 @@ and a 2-pass training procedure. The first pass includes unsupervised finetuning
|
|
| 11 |
For more details about LimaRP, see the model page for the [previously released version](https://huggingface.co/lemonilia/limarp-llama2-v2).
|
| 12 |
Most details written there apply for this version as well.
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
## Prompt format
|
| 15 |
Same as before. It uses the [extended Alpaca format](https://github.com/tatsu-lab/stanford_alpaca),
|
| 16 |
with `### Input:` immediately preceding user inputs and `### Response:` immediately preceding
|
|
|
|
| 11 |
For more details about LimaRP, see the model page for the [previously released version](https://huggingface.co/lemonilia/limarp-llama2-v2).
|
| 12 |
Most details written there apply for this version as well.
|
| 13 |
|
| 14 |
+
## Important notes on generation settings
|
| 15 |
+
It's recommended not to go overboard with low tail-free-sampling (TFS) values. From testing, decreasing it too much
|
| 16 |
+
appears to easily yield rather repetitive responses. Suggested starting generation settings are:
|
| 17 |
+
|
| 18 |
+
- TFS = 0.95
|
| 19 |
+
- Temperature = 0.70~0.85
|
| 20 |
+
- Repetition penalty = 1.05~1.10
|
| 21 |
+
- top-k = 0 (disabled)
|
| 22 |
+
- top-p = 1 (disabled)
|
| 23 |
+
|
| 24 |
## Prompt format
|
| 25 |
Same as before. It uses the [extended Alpaca format](https://github.com/tatsu-lab/stanford_alpaca),
|
| 26 |
with `### Input:` immediately preceding user inputs and `### Response:` immediately preceding
|