OwenArli commited on
Commit
cbc6182
1 Parent(s): 48ab4ab

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
+ ## RPMax Series Overview
8
+
9
+ | [2B](https://huggingface.co/ArliAI/Gemma-2-2B-ArliAI-RPMax-v1.1) |
10
+ [3.8B](https://huggingface.co/ArliAI/Phi-3.5-mini-3.8B-ArliAI-RPMax-v1.1) |
11
+ [8B](https://huggingface.co/ArliAI/Llama-3.1-8B-ArliAI-RPMax-v1.2) |
12
+ [9B](https://huggingface.co/ArliAI/Gemma-2-9B-ArliAI-RPMax-v1.1) |
13
+ [12B](https://huggingface.co/ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2) |
14
+ [20B](https://huggingface.co/ArliAI/InternLM2_5-20B-ArliAI-RPMax-v1.1) |
15
+ [22B](https://huggingface.co/ArliAI/Mistral-Small-22B-ArliAI-RPMax-v1.1) |
16
+ [70B](https://huggingface.co/ArliAI/Llama-3.1-70B-ArliAI-RPMax-v1.1) |
17
+
18
+ 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.
19
+
20
+ 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.
21
+
22
+ You can access the model at https://arliai.com and ask questions at https://www.reddit.com/r/ArliAI/
23
+
24
+ We also have a models ranking page at https://www.arliai.com/models-ranking
25
+
26
+ Ask questions in our new Discord Server! https://discord.gg/MedJbtdd
27
+
28
+ ## Model Description
29
+
30
+ ArliAI-RPMax-12B-v1.2 is a variant based on Mistral Nemo 12B Instruct 2407.
31
+
32
+ This is arguably the most successful RPMax model due to how Mistral is already very uncensored in the first place.
33
+
34
+ v1.2 update is a retrain using 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.
35
+
36
+ ### Training Details
37
+
38
+ * **Sequence Length**: 8192
39
+ * **Training Duration**: Approximately 2 days on 2x3090Ti
40
+ * **Epochs**: 1 epoch training for minimized repetition sickness
41
+ * **QLORA**: 64-rank 128-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/Mistral-Nemo-12B-ArliAI-RPMax-v1.2
50
+ * **GPTQ_Q4**: https://huggingface.co/ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2-GPTQ_Q4
51
+ * **GPTQ_Q8**: https://huggingface.co/ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2-GPTQ_Q8
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