trashpanda-org/Qwen2.5-72B-Azalea-v0

Recommended settings

Context/instruct template: ChatML.

Samplers: temperature at 0.9, min_p at 0.05, top_a at 0.3, TFS at 0.75, repetition_penalty at 1.03, DRY if you have access to it. Alternatively, top nsigma 1 with temp 1 worked fine during testing too.

Master import for ST: https://files.catbox.moe/b6nwbc.json

Thank you!

Big thanks to the folks in the trashpanda-org Discord server for testing and sending over some logs!

Reviews

Really liking the model, it's def up there with Snowdrop v0.

It's able to handle side chars really well, even between rerolls, feeling like it's not just a fluke that side chars are integrated.

Rerolls vary, which is good. Really loving the prose.

Thinking kinda weird since it's often hallucinating giving itself its own direction, but the end result still good, so I suppose it's not really a problem.

There's little to no slops, when there's some, it's really minor that I wouldn't really mind it.

It's yappy at times (sometimes you'd need more than 1k output), but I'd say even when it's yapping, it's a good yap.

Not as horny as in it'd jump on you, but it definitely teases. I'd say it's good, actually prefer this way.

There's no positivity bias for sure, which is a 👍

It's definitely smart, understanding my reply really well, at least in its final response, not sure wassup with some hallucination on the thinking as shown on the 4-5th images tho.

— Raihanbook

It can definitely get filthy, but it's impersonating {{user}} every once in a while.

Reasoning seems to work well? I'm currently just running off of the recommended sampler for Snowdrop and lowering the temp every few responses.

— Ema

Too much "heaven and earth" slop, impersonation from time to time. Good with NSFW in general but rushes. Long answers with good narrative, V3-like vibe, a bit positivity bias maybe? Rushes events.

Loved the fluff, sugar daddy smut was too soft. NSFW actions good but rushes to the end. Dead dove is unplayable. 6/10

— Carmenta

The model is unhinged and horny (a positive for me), it suffers issues from Qwen's quirk of being prone to impersonation like most Qwen 72B models amd still have some slops here and there.

However, that was under the circumstance of traditional samplers. idk, I haven't played with nsigma much, but I feel like it's refreshing. The prose improved so much and no impersonation across 10 swipes

I'll say with nsigma this model is 10/10 for me. But I'll want to use it on Featherless, which doesn't support such parameters. I am going to test a bit more with traditional samplers and leave more feedback while it is still up.

PS Tested this model after some claude/gemini log, so far no impersonation like my previous test did on a fresh chat

— OMGWTFBBQ

I've noticed that just like any LLMs would it seems to have a certain habit or sometimes repeating what it said, or rerolling but still it is similar to the previous responses on lower temp which is about 0.7-0.8 but on high temps such as 1.18 or 1 when its creative it seems to struggle with the consistency, which is length and context of response, not sure if that has anything to do with temps however.

As a non thinking model I am really impressed by its capabilities to generate responses that has high quality despite sometimes it started going onto its own merry way of repeating what it said (Not sure what the issue is but it does repeat pretty often even on high temp, not the full sentence but it'll be like probably the end part of a sentence will be the same as the first reroll?)

It follow prompts pretty well so that's a pos too!

It seems to not understand how to push the story onwards without the character leaving the scene if its response length was too long.

There is certain times when it's tweaking a bit but hey, what LLM won't tweak right :thumbsmile:

if Aza has a slightly better consistency in good quality responses I think it would be perfect, I enjoyed my roleplay with it otherwise!

— Sprout

ok so i've tried a few different bots with Azalea, and it's pretty freaking good. it stays really true to the characters and their personalities. as for impersonating - i haven't had that issue so far. i feel like Azalea is actually pretty great

from my further testing, it seems like it's either i get a perfect response within the first message, or i have to resend like 5 times to get another one of similar quality

— Shreksophoner

The first few responses were really good, but after a few rerolls the llm does seem to struggle with following a set structure. Content-wise, it was decent, left me some room to advance the story. Creativity wasn't anything mind-blowing, but i still think it did a decent job. Prose was creative too.

The quality of responses seems to be whats holding it back. Everything else was decent to good for me

— simon

You could say its like a wild schizophrenic man with nsigma. the consistency of responses is random; some of his stories are really interesting while other stories are just bat shit insane, but overall hes fun to be around

hes more boring on medication (no nsigma) so you convince him not to take it

— AIELO

Can maintain subtlety about things from nudges. I like it. Reasoning, even if it's not Snowdrop, keeps track of secrets and what char/user knows, it's sick. Comes up with nice little details now and again that's not in the card. Speech patterns take hold, character portrayal is on point most of the time, dialogue is decent to good. Prose is generally better than I remember Snowdrop had at times, worse in others, situational.

Without reasoning: It's fine, but narration seems to suffer compared to with-reasoning responses.

Writes more diverse exposition for smut than any Marigold ever did, interesting. I've noticed it can be more horny for thinking responses.

Drawbacks, however: tested it with a Spanish-speaking bot that Snowdrop does well with, and it's not interspersing Spanish in its responses like I'd expect it to. Other testers (and in my own testing), POV switch and user impersonation happens rarely. Tested with regular Qwen samplers for the most part - top nsigma didn't do well this time.

Overall, did pretty well in my testing.

— Severian

Some logs

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the TIES merge method using Columbidae/Qwen2.5-72B-Instruct as a base.

Models Merged

The following models were included in the merge:

• ./stage1
• ./stage2

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: ./stage2
    parameters:
      weight: 1
      density: 1
  - model: ./stage1
    parameters:
      weight: 1
      density: 1
merge_method: ties
base_model: Columbidae/Qwen2.5-72B-Instruct
parameters:
  normalize: true
  int8_mask: true
tokenizer_source: Columbidae/Qwen2.5-72B-Instruct
dtype: bfloat16
name: trashpanda-org/Qwen2.5-72B-Azalea-v0