Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,62 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
---
|
4 |
+
Qwen2.5-32B-ArliAI-RPMax-v1.3
|
5 |
+
|
6 |
+
=====================================
|
7 |
+
RPMax v1 Series Overview
|
8 |
+
|
9 |
+
v1.1 = 2B | 3.8B | 8B | 9B | 12B | 20B | 22B | 70B
|
10 |
+
|
11 |
+
v1.2 = 8B | 12B | 70B
|
12 |
+
|
13 |
+
v1.3 = 32B
|
14 |
+
|
15 |
+
RPMax is a series of models that are trained on a diverse set of curated creative writing and RP datasets with a focus on variety and deduplication. This model is designed to be highly creative and non-repetitive by making sure no two entries in the dataset have repeated characters or situations, which makes sure the model does not latch on to a certain personality and be capable of understanding and acting appropriately to any characters or situations.
|
16 |
+
|
17 |
+
Many RPMax users mentioned that these models does not feel like any other RP models, having a different writing style and generally doesn't feel in-bred.
|
18 |
+
|
19 |
+
You can access the model at https://arliai.com and we also have a models ranking page at https://www.arliai.com/models-ranking
|
20 |
+
|
21 |
+
Ask questions in our new Discord Server https://discord.com/invite/t75KbPgwhk or on our subreddit https://www.reddit.com/r/ArliAI/
|
22 |
+
Model Description
|
23 |
+
|
24 |
+
Qwen2.5-32B-ArliAI-RPMax-v1.3 is a variant made from the Qwen2.5-32B-Instruct model.
|
25 |
+
|
26 |
+
Let us know what you think of the model! The different parameter versions are based on different models, so they might all behave slightly differently in their own way.
|
27 |
+
|
28 |
+
v1.3 updated models are trained with updated software and configs such as the updated transformers library that fixes the gradient checkpointing bug which should help the model learn better.
|
29 |
+
This version also uses RSLORA+ for training which helps the model learn even better.
|
30 |
+
|
31 |
+
Specs
|
32 |
+
|
33 |
+
Context Length: 128K
|
34 |
+
Parameters: 32B
|
35 |
+
|
36 |
+
Training Details
|
37 |
+
|
38 |
+
Sequence Length: 8192
|
39 |
+
Training Duration: Approximately 4 days on 2x3090Ti
|
40 |
+
Epochs: 1 epoch training for minimized repetition sickness
|
41 |
+
LORA: 64-rank 64-alpha, resulting in ~2% trainable weights
|
42 |
+
Learning Rate: 0.00001
|
43 |
+
Gradient accumulation: Very low 32 for better learning.
|
44 |
+
|
45 |
+
Quantization
|
46 |
+
|
47 |
+
The model is available in quantized formats:
|
48 |
+
|
49 |
+
FP16: https://huggingface.co/ArliAI/Qwen2.5-32B-ArliAI-RPMax-v1.3
|
50 |
+
GGUF: https://huggingface.co/ArliAI/Qwen2.5-32B-ArliAI-RPMax-v1.3-GGUF
|
51 |
+
|
52 |
+
Suggested Prompt Format
|
53 |
+
|
54 |
+
ChatML Chat Format
|
55 |
+
|
56 |
+
<|im_start|>system
|
57 |
+
Provide some context and/or instructions to the model.
|
58 |
+
<|im_end|>
|
59 |
+
<|im_start|>user
|
60 |
+
The user’s message goes here
|
61 |
+
<|im_end|>
|
62 |
+
<|im_start|>assistant
|