ChuckMcSneed commited on
Commit
0b423ac
1 Parent(s): 99abda2

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +7 -24
README.md CHANGED
@@ -30,29 +30,12 @@ PS=creative writing
30
  Here they are. The results of each test. You can see pure data in file [LLM-test.csv](LLM-test.csv)
31
 
32
  What they show is quite interesting:
33
- - Goliath and Spicyboros are the best at following commands, followed by Qwen and Nous-Hermes
34
- - Qwen is terrible at creative writing, but good at following commands, Mixtral is the opposite
35
- - Xwin, Goliath, Spicyboros and Mixtral are the best at creative tasks
36
- - Una-xaberius shows that overtraining on benchmarks leads to loss of creativity and the model does not become smarter
37
- - Solar-instruct, despite its small size still can write poems, but is incapable of writing in style
38
- - ChatGPT can't pass B-test due to filter, C-test and P tests were not performed for that reason. It is incredibly good at stylized writing though, outperforming ALL tested local models. It can't pass D-test due to overfitting.
39
- - Spicyboros likely inherited its ability of stylized writing from GPT
40
- - Spicyboros is incredibly well finetuned on following instructions, and can even do stylized writing, outperforming Goliath on this benchmark. Not sure how it performs in practice
41
- - Cybertron seems to perform at approximately the same level as Solar-instruct, it also was surprisingly okay at writing poems
42
- - Neither Cybertron nor Solar-instruct outperform 70B models as they claim. Both are unable to follow advanced instructions(BCD tests).
43
- - Chronos is the first 70B model that can't pass any of BCD tests ie is incapable of following complex instructions
44
- - SynthIA is a disappointment
45
 
46
  # More tests?
47
- Feel free to suggest more models for testing by opening new discussion. Mention model name, size and why do you want to test it.
48
-
49
- # Updates
50
- ### 2023-12-22
51
- - Added Dolly15k, DiscoLM70B, SOLAR70B and qCammel
52
-
53
- ### 2023-12-20
54
- - Added Synthia, Spicyboros and Chronos
55
-
56
- ### 2023-12-19
57
- - Added solar-instruct(suspicously high benchmarks) una-xaberius(known cheater) and ChatGPT. Some tests were not performed with ChatGPT because Sam will ban me for them.
58
- - Added cybertron v3 per request of @fblgit.
 
30
  Here they are. The results of each test. You can see pure data in file [LLM-test.csv](LLM-test.csv)
31
 
32
  What they show is quite interesting:
33
+ - If a model can't pass any of the BCD tests, it is most likely braindead or very filtered(kinda same lol)
34
+ - If SP score of the model is very low it's writing style is dry
35
+ - Creative parent(Euryale) + creative parent(Xwin)=creative child(Goliath)
36
+ - Creative parent(Euryale) + dry parent(Nous-Hermes) + drier parent(SynthIA)=dry-ish child(Venus)
37
+ - Dry parent(Nous-Hermes) + creative parent(Xwin) + creative parent(Mythospice)=creative child(lzlv)
38
+ - Cheater meme model(una-cybertron) was somewhat creative, but braindead
 
 
 
 
 
 
39
 
40
  # More tests?
41
+ Feel free to suggest more models for testing by opening new discussion. Mention model name, size and why do you want to test it.