Qwen2.5-32B-Marigold-v1

Iterating on Marigold in hopes of another Snowdrop?

Severian's notes: Not going to lie, this model candidate won when I was honestly preferring another one. Doesn't help that we're all coming off a collective Snowdrop high. Most of the model testers liked this over the other candidates as well as v0, but it's not drastic in terms of improvements. A little sloppy, thinking holds better here than v0. The character portrayal is stronger with thinking on. It's slightly unhinged, too.

Has's notes: honestly this one kinda sucks, but then again it was just Mullein v1 data mixture on top of Qwen 32B, so I'm not sure why I was expecting something special

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.

A virt-io derivative prompt worked best during our testing, but feel free to use what you like. Seems to perform decent (and better than otherwise) with thinking.

60+ followers!

Thanks for the support, y'all.

Thank you!

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

Reviews

PROS:

Most proactive char stance. Even if char himself wasn’t particularly doing anything, their behaviour and inner monologue showed a clear directive and their current goal.

Kept char’s attitude consistent. Also portrayed inner conflict well between duty and personal attachment.

All swipes gave very different outputs. Positive or negative, those were still very different from one another. Definitely most interesting prose, too.

One of the swipes went into action by itself and it was written out great. Stakes were high, sense of danger was present.

CONS:

Jumped to romantic tension a tad bit too quickly. Literally second message and I got hit with a ‘his fingers brushed against their neck’. I’m about to get (possibly) executed and he’s trying to flirt.

Had just some minor odd capitalization like ‘she kNEW’.

Conclusion: Definitely my favorite out of them all. Speaking for user was close to none, every response was very vibrant and kept me engaged (to the point where I actually want to continue rp-ing with it out of personal preference). Just very good variety in responses. Barely spoke for user, kept everything in character. Mwuah!

— Sellvene

Non-thinking: Best responses so far [out of Marigold v1 candidates]. Not extraordinary but good responses for [being a] small-mid model.

Thinking: Seems like would work nice in smut and felt best among models. But still, seemed a bit too basic for my R1-Snowdrop addicted ass.

— Carmenta

Gacha with rare responses out of 1/4 rolls

— AIELO

I was trying qwq the other day and I keep running into impersonation issue, I was wondering if this because I'm using base model and it wasn't trained with roleplaying data. It thinks, but the issue persist. Y'know, talking and acting as user.

On the contrary, thinking curbs that completely for marigold, it reinforces the model to stick as the character, ignoring user completely. There's still some phrases that describe what the user is doing, but it doesn't get expanded on. There's a small issue with not getting added though, so thinking and actual response doesn't get separated.

The prose, it has some slop here and there, but it's pretty tolerateable honestly.

— Sam

It's able to impersonate side chars, and the fact that they're interacting is the better. I feel like it repeats response structure and the slops are juuuust...

— Raihanbook