Surpised `Eurux-8x22b-nca` did so well!?

#4
by jukofyork - opened

I've just been experimenting with these models and found this merge to be better at coding:

models:
  - model: Mixtral-8x22B-v0.1
  - model: Eurux-8x22b-kto
  - model: Eurux-8x22b-nca
base_model: Mixtral-8x22B-v0.1
merge_method: model_stock
dtype: float16

As is this merge:

models:
  - model: codellama/CodeLlama-70b-hf
  - model: jukofyork/Eurus-70b-sft-fixed
  - model: jukofyork/Eurus-70b-nca-fixed
base_model: codellama/CodeLlama-70b-hf
merge_method: model_stock
dtype: float16

(this actually has lower perplexity too!)

Don't use Eurus-70b-merge for anything creative though... It can't write stories at all and merging it into miqu-1 as I hoped (ie: it's the only other model with 1M base RoPE and uses the same [INST] template) just makes the merge equally bad... :/


They are both very finicky about the prompt template though:

<s> [INST] {Prompt} [/INST] {Response}

Not having a space after the <s> or [/INST] degrades their coding ability really badly...

What prompt template did you use? I just tested Eurux-8x22b-merged and it writes horrible/bad stories for me?! Maybe another prompt template works better for creative stuff?

EDIT: Maybe not 100% horrible/bad, but every sentence is very short and the language used isn't very descriptive...

Yeah:

<s>[INST] {Prompt} [/INST]{Response}

and:

<s> [INST] {Prompt} [/INST]{Response}

both seem to work better for stories. Even the Eurus-70b-merge isn't that bad now :/

Now I'm going to have to redo all the miqu-1 merge tests to see what happens...

Surpised Eurux-8x22b-nca did so well!?

If you look at SP("creative tasks") score, it didn't. Wizard-8x22b and Mixtral-8x22B-Instruct outperformed it. It was quite obedient though(high BCD).

What prompt template did you use? I just tested Eurux-8x22b-merged and it writes horrible/bad stories for me?! Maybe another prompt template works better for creative stuff?

Don't remember, probably Alpaca.

Surpised Eurux-8x22b-nca did so well!?

If you look at SP("creative tasks") score, it didn't. Wizard-8x22b and Mixtral-8x22B-Instruct outperformed it. It was quite obedient though(high BCD).

What prompt template did you use? I just tested Eurux-8x22b-merged and it writes horrible/bad stories for me?! Maybe another prompt template works better for creative stuff?

Don't remember, probably Alpaca.

Ah, I think I've had my fill of "[INST]" prompts... Seems like every model that uses this just ends up a finicky PITA to get working properly... :(

Ah, I think I've had my fill of "[INST]" prompts... Seems like every model that uses this just ends up a finicky PITA to get working properly... :(

I have this issue as well. Good thing is, they usually understand Alpaca for some reason. Maybe because [INST] and ### Instruction are close in vector space?

Yeah, ever since WolfRam did his prompt format comparison test on 8x7B Instruct I have been converted to Alpaca on [INST] Mistral models

Yeah, I think models like "###" because of seeing a lot of markdown data too. Coding models definitely seem to work better if you put "# " before your task - I think this is because they have been massively over trained on Python and "# Your task is" is like a lot of what they see during training.

Yeah, ever since WolfRam did his prompt format comparison test on 8x7B Instruct I have been converted to Alpaca on [INST] Mistral models

The weird thing is Mistral, Mixtral and Miqu all seemed to want a slightly different way of adding spaces and even the official pages and tokenizer_configfiles didn't always agree.

I can't get Mixtral-8x22B-Instruct to work whatever I do lol. I've tried every possible combination of the prompt template and even ran on CPU at fp16 just to see if it was the quantitation causing problems, but it basically works like a broken frankenmerge for me - randomly stopping all the time mid-sentence, etc.

I think the GGUF conversation is going wrong with the extra vocab they added for the instruct model or something... :(

I have an old version of Mixtral-8x22B-Instruct quant and it works just fine with Alpaca, quality is nothing special though. Must be the new update that messes it up.

image.png
Seems like they really overfitted it compared to Mistral-medium https://arxiv.org/abs/2405.00332, that may explain why I still prefer Miqu over those new 8x22b models, even when they perform better on benches.

Sign up or log in to comment