trashpanda-org/QwQ-32B-Snowdrop-v0.5-Type-R

One of three experimental merges in an attempt to make Snowdrop v1.

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.

A virt-io derivative prompt worked best during our testing, but feel free to use what you like.

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

While it doesn’t express emotions as effectively as Snowdrop v0, its responses were longer and more creative, but jumped between formatting.

Compared to v0, this model focused more on the topic, handling it well while staying true to char personality. ... This response feels like R1 in a good way. It stays true to his core personality and handles the issue appropriately, even referencing that this has happened before, which is a plus. ... It included details that the previous model didn’t mention. It’s creative, introducing another NPC mentioned in the intro and personality.

— Azriael

I found it interesting how different Type-R seems to focus on different aspects or specifically weight a certain aspect more somehow? Or rather like Type-S where its reasoning seems to lean into the emotional aspect more, it's just something I noticed.

But Type-R do have a more consistent reasoning comparing with Type-S, it's just the response it generate was slightly off quality, for example: "Suddenly hyper-aware" out of nowhere and wrapping reality when the char only have gravitational power..maybe the temp was too high? But again, just like Vennie, creativity is appreciated but it can try harder next time.

— Sprout

SMUT: seems to be more focused on char's inner monologue and their feelings rather than describing the scene/movement itself. 2/5 swipes were almost identical to one another, however. Also seemed to want to rush to the end as quickly as possible.

PROS:

Steady in its writing. You won't get random hiccups of Chinese or anything, the style seems to be consistent too.

Keeps characterization grounded, but in line with personality provided in the card. Also utilizes rp time period accordingly the most.

Doesn't seem to be showing any positivity bias.

Creative when describing the atmosphere, especially if you write a response on the longer side.

Good prose, decent dialogue immersion. Rarely went OOC route.

CONS:

Inconsistent formatting. Sometimes direct speech gets wrapped in asterisks or additional quotations, sometimes it doesn't. Not dependent on how user writes/first message, will just do it randomly out of nowhere. Also doesn't mark down quoting user's words in any way, unlike other models.

A bit too bland and repetitive. Nothing too bad or anything, just... too safe for its own good. Similar swipes across the board.

Some major cases of talking for user, although easily fixable. Probably the only model that consistently spoke for user across all bots I've tested it on.

Conclusion: It is a steady model, feels almost identical to v0. While it's great, it's not doing anything new. I'd rather take risks with a newer, more creative model, than play it safe with some minor almost 'bug fixes' that this model seems to be. It's good if you want more story-driven slow burn (enemies to lovers/modern day setting especially from my tests), but anything too out of regular scope and it kind of tanks in creativity.

— Sellvene

I'm a bit indecisive about this one, character cohesion is.. off. Sometimes mixing who is who, I'm not getting impersonation issue though, might due to thinking again.

Formatting is inconsistent across reroll, and sometimes it repeats your input by adding ? At the end. Or "she repeats"

And rerolls... how do I put this. Pretty similar, but they take the context from the first message well.

But hey, no positivity bias for angst bot. Reasoning pretty spot on with this character core personality traits. It's amazing.

— Sam

Horniness: 4/5 After seeing this is [the prior test model name], it made me wonder what happened. I remember liking it but this seems hit or miss to me. It can definitely cook, but feel inconsistent quality and formatting wise.

— BBQ

[PDF omitted for brevity] This model scored second place with an RP score of 65% compared to Snowdrop v0 with an RP score of 58%, especially with characterization and creativity on both text completion presets.

Cohesion and responsiveness remained similar throughout all (experimental) models. Responses consistently included one or more paragraphs that were varied in the ratio of speech and narration, ranging from 10% to 60% character dialogue to narration. Cohesion remained intact throughout most of the trials, with the style of narration of each model being similar to each other. Results with top nsigma use were significantly better.

— AIELO

Not bad but after Snowdrop v0... it has slops :/

— Carmenta

It's able to handle side chars. Well, at least to mention and incorporate them into the scene with little dialogues

Rerolls feel different from each other, which is good

I somehow feel like the responses sort of degraded from the first message, not as much description, not sure why. Tried rerolling, didn't seem to help

Model isn't too horny, which is good in my book

I feel like it's tad bit positive in my opinion, tho it's subjective

— Raihanbook

Using top nsigma, every response in my test followed the same train of thought, generally following the same format making it pretty stale. Not good at all, even if reasoning and dialogue end up decent in average. Regular Qwen samplers oddly make it more creative across rerolls, strangely enough.

I prefer Type-S and Type-H over this, tbh.

— 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/Qwen25-32B as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: trashpanda-org/Qwen2.5-32B-Marigold-v1
    parameters:
      weight: 1
      density: 1
  - model: trashpanda-org/Qwen2.5-32B-Marigold-v0
    parameters:
      weight: 1
      density: 1
  - model: Columbidae/QwQ-32B
    parameters:
      weight: 0.9
      density: 0.9
merge_method: ties
base_model: Columbidae/Qwen25-32B
parameters:
  weight: 0.9
  density: 0.9
  normalize: true
  int8_mask: true
tokenizer_source: Columbidae/Qwen25-32B-Instruct
dtype: bfloat16

trashpanda-org
/

QwQ-32B-Snowdrop-v0.5-Type-R