concedo commited on
Commit
0f37e35
1 Parent(s): 79f2786

Updated Readme.md

Browse files
Files changed (1) hide show
  1. README.md +48 -0
README.md CHANGED
@@ -1,3 +1,51 @@
1
  ---
2
  license: other
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: other
3
+ language:
4
+ - en
5
+ inference: false
6
  ---
7
+
8
+ # OPT-2.7B-Nerybus-Mix
9
+
10
+ This is an experimental model containing a ***parameter-wise 50/50 blend (weighted average)*** of the weights of *NerysV2-2.7B* and *ErebusV1-2.7B*
11
+ Preliminary testing produces pretty coherent outputs, it appears to retain the NSFWness of Erebus but with a Nerys-esque twist in terms of prose.
12
+
13
+ # License
14
+ The two models used for this blend, *NerysV2-2.7B* and *ErebusV1-2.7B* are made by **Mr. Seeker**.
15
+ - https://huggingface.co/KoboldAI/OPT-2.7B-Erebus
16
+ - https://huggingface.co/KoboldAI/OPT-2.7B-Nerys-v2
17
+ The base OPT-2.7B model is licensed under the OPT-175B license, Copyright (c) Meta Platforms, Inc. All Rights Reserved.
18
+
19
+ # Evaluation Results
20
+ As the original datasets used for the source models are not publically available, I use my own datasets for this evaluation, which may not provide accurate comparison.
21
+
22
+ Eval parameters: 25000 characters extracted from the middle of the corpus, tested in blocks of 1024 tokens each, same dataset used for each test batch.
23
+
24
+ ```
25
+ Literotica Dataset Eval (Randomly selected stories)
26
+ {'eval_loss': 2.592170000076294, 'name': 'Concedo_OPT-2.7B-Nerybus-Mix'}
27
+ {'eval_loss': 2.571096181869507, 'name': 'KoboldAI_OPT-2.7B-Erebus'}
28
+ {'eval_loss': 2.6392364501953125, 'name': 'KoboldAI_OPT-2.7B-Nerys'}
29
+ {'eval_loss': 2.4650306701660156, 'name': '(Unreleased 2.7B ModronAI Model)'}
30
+
31
+ ASSTR Dataset Eval (Randomly selected stories)
32
+ {'eval_loss': 2.604311227798462, 'name': 'Concedo_OPT-2.7B-Nerybus-Mix'}
33
+ {'eval_loss': 2.5843987464904785, 'name': 'KoboldAI_OPT-2.7B-Erebus'}
34
+ {'eval_loss': 2.666102170944214, 'name': 'KoboldAI_OPT-2.7B-Nerys'}
35
+ {'eval_loss': 2.9165072441101074, 'name': '(Unreleased 2.7B ModronAI Model)'}
36
+
37
+ Harry Potter Dataset Eval (Canon books)
38
+ {'eval_loss': 2.391289234161377, 'name': 'Concedo_OPT-2.7B-Nerybus-Mix'}
39
+ {'eval_loss': 2.40213680267334, 'name': 'KoboldAI_OPT-2.7B-Erebus'}
40
+ {'eval_loss': 2.4142935276031494, 'name': 'KoboldAI_OPT-2.7B-Nerys'}
41
+ {'eval_loss': 2.227642774581909, 'name': '(Unreleased 2.7B ModronAI Model)'}
42
+
43
+ Star Wars Dataset Eval (Rogue One Novel)
44
+ {'eval_loss': 2.4152939319610596, 'name': 'Concedo_OPT-2.7B-Nerybus-Mix'}
45
+ {'eval_loss': 2.4259495735168457, 'name': 'KoboldAI_OPT-2.7B-Erebus'}
46
+ {'eval_loss': 2.449702024459839, 'name': 'KoboldAI_OPT-2.7B-Nerys'}
47
+ {'eval_loss': 2.788408041000366, 'name': '(Unreleased 2.7B ModronAI Model)'}
48
+
49
+ ```
50
+
51
+ It is recommend to use this model with the KoboldAI software. All feedback and comments can be directed to Concedo on the KoboldAI discord.