OwenArli commited on
Commit
63e1275
1 Parent(s): 8019936

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +57 -3
README.md CHANGED
@@ -1,3 +1,57 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ ---
4
+ # ArliAI-RPMax-12B-v1.2
5
+ =====================================
6
+
7
+ ## UPDATE: Merged it wrongly to base after LORA training. Reuploaded all the files and working on reuploading the GGUF. The 8B version meanwhile is working fine.
8
+
9
+ ## RPMax Series Overview
10
+
11
+ | [2B](https://huggingface.co/ArliAI/Gemma-2-2B-ArliAI-RPMax-v1.1) |
12
+ [3.8B](https://huggingface.co/ArliAI/Phi-3.5-mini-3.8B-ArliAI-RPMax-v1.1) |
13
+ [8B](https://huggingface.co/ArliAI/Llama-3.1-8B-ArliAI-RPMax-v1.2) |
14
+ [9B](https://huggingface.co/ArliAI/Gemma-2-9B-ArliAI-RPMax-v1.1) |
15
+ [12B](https://huggingface.co/ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2) |
16
+ [20B](https://huggingface.co/ArliAI/InternLM2_5-20B-ArliAI-RPMax-v1.1) |
17
+ [22B](https://huggingface.co/ArliAI/Mistral-Small-22B-ArliAI-RPMax-v1.1) |
18
+ [70B](https://huggingface.co/ArliAI/Llama-3.1-70B-ArliAI-RPMax-v1.1) |
19
+
20
+ 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.
21
+
22
+ Early tests by users mentioned that these models does not feel like any other RP models, having a different style and generally doesn't feel in-bred.
23
+
24
+ You can access the model at https://arliai.com and ask questions at https://www.reddit.com/r/ArliAI/
25
+
26
+ We also have a models ranking page at https://www.arliai.com/models-ranking
27
+
28
+ Ask questions in our new Discord Server! https://discord.com/invite/t75KbPgwhk
29
+
30
+ ## Model Description
31
+
32
+ ArliAI-RPMax-12B-v1.2 is a variant based on Mistral Nemo 12B Instruct 2407.
33
+
34
+ This is arguably the most successful RPMax model due to how Mistral is already very uncensored in the first place.
35
+
36
+ v1.2 update completely removes non-creative/RP examples in the dataset and is also an incremental improvement of the RPMax dataset which dedups the dataset even more and better filtering to cutout irrelevant description text that came from card sharing sites.
37
+
38
+ ### Training Details
39
+
40
+ * **Sequence Length**: 8192
41
+ * **Training Duration**: Approximately 2 days on 2x3090Ti
42
+ * **Epochs**: 1 epoch training for minimized repetition sickness
43
+ * **LORA**: 64-rank 128-alpha, resulting in ~2% trainable weights
44
+ * **Learning Rate**: 0.00001
45
+ * **Gradient accumulation**: Very low 32 for better learning.
46
+
47
+ ## Quantization
48
+
49
+ The model is available in quantized formats:
50
+
51
+ * **FP16**: https://huggingface.co/ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2
52
+ * **GGUF**: https://huggingface.co/ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2-GGUF
53
+
54
+
55
+ ## Suggested Prompt Format
56
+
57
+ Mistral Instruct Prompt Format