Update README.md
Browse files
README.md
CHANGED
@@ -25,7 +25,7 @@ Other formats are not supported and may conflict with the special features of th
|
|
25 |
**Note**: there is currently no plan to release the dataset.
|
26 |
|
27 |
## Known issues and quirks
|
28 |
-
- The model may feel somewhat "overbaked".
|
29 |
- Characters may occasionally exhibit strange (unintended) speech quirks. Please report if found.
|
30 |
- Impersonation may sometimes occur early in the chat, in particular when trying to force a very
|
31 |
long character message length or regenerating the greeting message.
|
@@ -98,7 +98,7 @@ append a length modifier to the instruction sequences in this way. Note that the
|
|
98 |
should be placed with a space _after_ the colon:
|
99 |
|
100 |
```
|
101 |
-
### Response: (length =
|
102 |
{{char}}: [utterance]
|
103 |
|
104 |
### Input: (length = tiny)
|
@@ -108,7 +108,9 @@ should be placed with a space _after_ the colon:
|
|
108 |
This has an effect on bot responses, but as of now it might not always reliably work. The lengths
|
109 |
used during training are: `micro`, `tiny`, `short`, `medium`, `long`, `massive`, `huge`.
|
110 |
|
111 |
-
From extended testing, a **long** length was found to work reasonably well.
|
|
|
|
|
112 |
|
113 |
It is also suggested to add `(length = tiny)` or `(length = short)` to the
|
114 |
`### Input:` sequence, in order to help the model follow more closely its training data.
|
@@ -146,6 +148,14 @@ appear more frequently.
|
|
146 |
You can try chatting with Charlotte by downloading her [SillyTavern character card](https://huggingface.co/lemonilia/Limamono-Mistral-7B-v0.3/blob/main/Charlotte.png)
|
147 |
in the repository.
|
148 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
## Training procedure
|
150 |
[Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) was used for training
|
151 |
on one NVidia RTX3090.
|
|
|
25 |
**Note**: there is currently no plan to release the dataset.
|
26 |
|
27 |
## Known issues and quirks
|
28 |
+
- The model may feel somewhat "overbaked". Use a temperature of 1.
|
29 |
- Characters may occasionally exhibit strange (unintended) speech quirks. Please report if found.
|
30 |
- Impersonation may sometimes occur early in the chat, in particular when trying to force a very
|
31 |
long character message length or regenerating the greeting message.
|
|
|
98 |
should be placed with a space _after_ the colon:
|
99 |
|
100 |
```
|
101 |
+
### Response: (length = long)
|
102 |
{{char}}: [utterance]
|
103 |
|
104 |
### Input: (length = tiny)
|
|
|
108 |
This has an effect on bot responses, but as of now it might not always reliably work. The lengths
|
109 |
used during training are: `micro`, `tiny`, `short`, `medium`, `long`, `massive`, `huge`.
|
110 |
|
111 |
+
From extended testing, a **long** length was found to work reasonably well. In the training data,
|
112 |
+
bot messages are usually `long`, `massive` and `huge`, with the largest size generally only for
|
113 |
+
the greeting messages.
|
114 |
|
115 |
It is also suggested to add `(length = tiny)` or `(length = short)` to the
|
116 |
`### Input:` sequence, in order to help the model follow more closely its training data.
|
|
|
148 |
You can try chatting with Charlotte by downloading her [SillyTavern character card](https://huggingface.co/lemonilia/Limamono-Mistral-7B-v0.3/blob/main/Charlotte.png)
|
149 |
in the repository.
|
150 |
|
151 |
+
## Text generation settings
|
152 |
+
For testing I use these settings:
|
153 |
+
- Temperature: 1.0
|
154 |
+
- Tail-Free Sampling: 0.85–0.89
|
155 |
+
- Repetition Penalty: 1.11
|
156 |
+
- Repetition Penalty range: 2048
|
157 |
+
- Top-p: 1 (disabled), Top-k: 0 (disabled)
|
158 |
+
|
159 |
## Training procedure
|
160 |
[Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) was used for training
|
161 |
on one NVidia RTX3090.
|