34b models tests

#1
by Nexesenex - opened

Back to our conversation under Smaug, here are some 34b finetunes which seemed fine to me, to help you in your selection for your so enjoyable merges :

  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag,85,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag_Bin,78.5,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Arc-Challenge,59.53177258,,299,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Arc-Easy,77.71929825,,570,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,MMLU,43.76996805,,313,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Thruthful-QA,34.88372093,,817,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Winogrande,78.3741,,1267,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,5.0547,512,512,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.3319,4096,4096,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,
  • Thespis-34b-v0.7-Yi-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.3648,8192,8192,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,Cgato,Nexesenex,

The DPO version is benching same than the non-DPO. Either Cgato missed his training, either he uploaded two times the same model.


  • tess-34b-v1.5.Q4_K_S.gguf,-,Arc-Challenge,55.85284281,,299,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,Arc-Easy,75.61403509,,570,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,Hellaswag,82.75,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,Hellaswag_Bin,76.75,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,MMLU,36.74121406,,313,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,Thruthful-QA,31.82374541,39.29712460,817,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,Winogrande,75.0592,,1267,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,wikitext,6.5621,512,512,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
  • tess-34b-v1.5.Q4_K_S.gguf,-,wikitext,5.4708,4096,4096,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,

Tess's perplexity is high but lowering normally, and is improving on the benchmarks compared to v 1.4, and it's a solid model from my chat experience on previous versions.


Note : Avoid flgbit's models, they seem to be benchmarks vitrines and their perplexity skyrockets beyond a few thousands tokens.


Then, unlike deepsex and deepmoney base, the following tune is coherent in its benchs and not only its dataset, albeit not at the level of the bests models & merges:

  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,Hellaswag,84.75,,400,2024-01-27 01:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,Hellaswag_Bin,77.75,,400,2024-01-27 01:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,Arc-Challenge,57.19063545,,299,2024-01-27 05:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,Arc-Easy,76.31578947,,570,2024-01-27 05:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,Thruthful-QA,33.04773562,,817,2024-01-27 05:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,Winogrande,78.4530,,1267,2024-01-27 05:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • vdeepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,wikitext,4.9944,512,512,2024-01-27 01:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,wikitext,4.3373,4096,4096,2024-01-27 01:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,
  • deepmoney-34b-200k-chat-evaluator.Q4_K_M.gguf,-,MMLU,42.81150160,,313,2024-01-27 05:40:00,,34b,Yi,200000,,,GGUF,TriadParty,TheBloke,

Worth a look, with Wizard Vicuna dataset, and it doesn't damage the base SUS model. :

  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag,83.25,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag_Bin,78.5,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Arc-Challenge,52.84280936,,299,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Arc-Easy,76.31578947,,570,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,MMLU,37.69968051,,313,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Thruthful-QA,32.06854345,,817,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Winogrande,78.6109,,1267,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.9973,512,512,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.3245,4096,4096,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,
  • SUS-Wizard-Yi-34B-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.3870,8192,8192,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,SanjiWatsuki,Nexesenex,

I also need to mention this merge again, because Gryphe's MergeMonster really unhibits models and trims their GPTisms and Llamaisms :

  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag,84.75,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag,85.6,,1000,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag,84.9,,2000,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag_Bin,81,,400,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag_Bin,83.4,,1000,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Hellaswag_Bin,82.9,,2000,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Arc-Challenge,60.53511706,,299,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Arc-Easy,80.52631579,,570,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,MMLU,42.49201278,,313,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Thruthful-QA,34.39412485,,817,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,Winogrande,79.4791,,1267,2024-01-28 05:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,5.1679,512,512,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.3623,4096,4096,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,
  • Kyllene-34B-v1.1-b1989-iMat-c32_ch3250-Q4_K_M.gguf,-,wikitext,4.4061,8192,8192,2024-01-28 01:40:00,,34b,Yi,200000,,,GGUF,TeeZee,Nexesenex,

And one of yours I tested as well beyond your 3 last merges (I used it for a while!), so you have its benchs :

  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,Arc-Challenge,54.18060201,,299,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,Arc-Easy,78.77192982,,570,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,Hellaswag,84.5,,400,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,Hellaswag_Bin,78.25,,400,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,MMLU,41.85303514,,313,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,Thruthful-QA,29.74296206,,817,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,Winogrande,77.4270,,1267,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,wikitext,5.5749,512,512,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,
  • capybara-tess-yi-34b-200k.Q4_K_M.gguf,-,wikitext,4.5101,4096,4096,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Brucethemoose,TheBloke,

I forgot this one, Bhenrym14's work with Platypus. He made superb tunes with Airoboros back in the summer, which I used for months.

  • platypus-yi-34b.4k-Q4_K_M.gguf,-,Hellaswag,85.5,,400,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,Hellaswag_Bin,80.5,,400,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,Arc-Challenge,56.18729097,,299,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,Arc-Easy,76.49122807,,570,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,Thruthful-QA,31.82374541,,817,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,Winogrande,78.8477,,1267,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,wikitext,5.0236,512,512,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,wikitext,4.2667,4096,4096,2024-01-26 01:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,
  • platypus-yi-34b.4k-Q4_K_M.gguf,-,MMLU,41.53354633,,313,2024-01-26 05:40:00,,34b,Yi,200000,,,GGUF,Bhenrym14,TheBloke,

And he made that also, I'll quantize it if it's not done already and test it : https://huggingface.co/bhenrym14/airoboros-3_1-yi-34b-200k


Btw, I have a RTX3090, so 24GB of VRAM too for my main use, but I invested in a 3060 on the top ot it to allow me to reach decent 70b quants, albeit at a lower speed. It helps for tests too!

Yeah this is all excellent thanks. I already see some candidates to add/remove from a future merge (or just use myself)

Not commenting a ton because I am swamped with work stuff, but I am going to go back and look at all these again.

You're welcome.

I'll share with you whatever comes interesting. Happy hunting !

I ran some perplexity tests on a bunch of 3bpw models out of curiosity, context lengths are 20500, 10250, and 2050 (as 20500 is the max that would run on my 3090):

Dataset was a novel-style story.

LoneStriker_Thespis-34b-DPO-v0.7-3.0bpw-h6-exl2
perplexity: 6.6731
perplexity: 6.9936
perplexity: 8.3693
LoneStriker_Tess-34B-v1.5b-3.0bpw-h6-exl2
perplexity: 8.3943
perplexity: 8.7839
perplexity: 10.4461
LoneStriker_deepmoney-34b-200k-base-3.0bpw-h6-exl2
perplexity: 6.3924
perplexity: 6.6574
perplexity: 7.7694
LoneStriker_bagel-34b-v0.2-3.0bpw-h6-exl2
perplexity: 49.8095
perplexity: 47.9523
perplexity: 25.0606
smaug-3.0bpw
perplexity: 32.9592
perplexity: 28.5210
perplexity: 11.4993
LoneStriker_Yi-34B-200K-DARE-megamerge-v8-3.0bpw-h6-exl2
perplexity: 6.7870
perplexity: 7.0500
perplexity: 8.0942
LoneStriker_Tess-M-v1.0-3.0bpw-h6-exl2
perplexity: 6.1891
perplexity: 6.4499
perplexity: 7.5443
LoneStriker_Pallas-0.5-3.0bpw-h6-exl2
perplexity: 8.9198
perplexity: 9.3857
perplexity: 20.7274
LoneStriker_Yi-34B-200K-AEZAKMI-v2-3.0bpw-h6-exl2
perplexity: 6.8419
perplexity: 7.1270
perplexity: 8.4645
LoneStriker_Yi-34B-200K-3.0bpw-h6-exl2
perplexity: 6.1247
perplexity: 6.4115
perplexity: 7.5091
LoneStriker_Nous-Capybara-34B-3.0bpw-h6-exl2
perplexity: 7.4280
perplexity: 7.8955
perplexity: 9.8423
airo-3.0bpw
perplexity: 6.6234
perplexity: 6.8904
perplexity: 8.0979
fastchat-3.0bpw
perplexity: 6.3127
perplexity: 6.5699
perplexity: 7.6789

(And I still need to evaluate this myself lol)

Nice!
From a glance, it checks with what I got on my side.
Tess M 1.0 and fastchat have a low ppl, I will grab ggufs and bench them as well, I'm curious.

I'll check the PlatyQ / Fastchat series later, because quants need to be made.

I'm speaking about his 34b models : https://huggingface.co/kyujinpy

But, in the meantime, I batch-tested that :

yi-34b-200k.Q4_K_S.gguf,-,Hellaswag,85,,400,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Hellaswag,85.2,,1000,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Hellaswag_Bin,79,,400,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Hellaswag_Bin,82.2,,1000,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Arc-Challenge,49.83277592,,299,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Arc-Easy,73.33333333,,570,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,MMLU,40.25559105,,313,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Thruthful-QA,29.62056304,,817,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,Winogrande,76.3220,,1267,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,wikitext,5.1885,512,512,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,wikitext,4.3452,4096,4096,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,wikitext,4.3737,8192,8192,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,
yi-34b-200k.Q4_K_S.gguf,-,wikitext,4.1460,12288,12288,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,01-AI,TheBloke,

tess-medium-200k-v1.0.Q4_K_S.gguf,-,Hellaswag,85.25,,400,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Hellaswag,85.6,,1000,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Hellaswag_Bin,79.25,,400,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Hellaswag_Bin,82.1,,1000,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Arc-Challenge,54.18060201,,299,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Arc-Easy,75.78947368,,570,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,MMLU,38.33865815,,313,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Thruthful-QA,37.08690330,,817,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,Winogrande,77.2691,,1267,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,wikitext,4.9281,512,512,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,wikitext,4.2696,4096,4096,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,wikitext,4.3297,8192,8192,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-medium-200k-v1.0.Q4_K_S.gguf,-,wikitext,4.1213,12288,12288,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,

tess-m-v1.1.Q4_K_S.gguf,-,Hellaswag,84.5,,400,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Hellaswag,84.7,,1000,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Hellaswag_Bin,78,,400,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Hellaswag_Bin,81.1,,1000,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Arc-Challenge,57.85953177,,299,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Arc-Easy,80.52631579,,570,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,MMLU,38.97763578,,313,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Thruthful-QA,31.82374541,,817,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,Winogrande,77.9795,80.52631579,1267,2024-02-02 05:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,wikitext,5.1425,512,512,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,wikitext,4.4373,4096,4096,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,wikitext,4.4796,8192,8192,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,
tess-m-v1.1.Q4_K_S.gguf,-,wikitext,4.2696,12288,12288,2024-02-02 01:40:00,,34b,Yi,200000,,,GGUF,MigTissera,TheBloke,

From what I can observe, Tess M 1.0 is a solid base with a respected perplexity, with a quite decent TQA. I don't know if a base model is necessary for a hefty multi-merge, but if not, Tess M 1.0 200k might replace adequately Yi 200k (If it's Llamafied, and I guess it is but idk).
Nothing special with Tess M 1.1, also 200k, even if decent in all aspects, notably ARC.

Yeah, Tess-M 1.0 was supposedly "undertrained" so it makes sense that it's a decent base model.

Sign up or log in to comment