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

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

Narration was cut clean and simple. It listened to the character card extremely well, lorebook W's, and also added some creativity. The CoT was very simple, too. The dialogue was great for me and I really enjoyed it!

— Mooth

It’s so creative that it often ignores parts of the personality just to create something different with each swipe.

— Azriael

PROS:

Threw some absolute banger lines my way. Even during smut it managed to come up with some lines that made me laugh out loud.

Actually made me read reasoning for once, lol. Probably most unique one of them all in that department

Incredibly varied swipes. I may or may not have gotten carried away and did 7 swipes instead of 5 just because of how interesting it was to read.

Great at utilizing creativity as a way to fill up world building. Some things may seem like a bit too much, but I'd rather take too much than none at all.

Actually kept great NPC characterisation. No major mishaps, kept them as a part of the current scene instead of making them into living props.

Better spatial awareness than most models on this weekends rotation.

Awesome prose. Stuck to theme, setting and char's speech pattern perfectly most of the time. Least slop I've seen.

CONS:

A stubborn bastard. Even OOC prompting sometimes didn't help. If it wants to speak for user it WILL.

Odd formatting. (E.G screenshots 3&6.) Wouldn't break out of the pattern even with a gun held to its head.

Can be a bit too creative for it's own good, jumping over its head, which will end up ruining the roleplay pacing. Struggles with unique anatomy (horns, wings, etc) as well.

Loves to do unnecessary time skips/4th wall breaks. A bit too unhinged for its own good.

Conclusion: Probably the best one from a creative standpoint. One major downside - unpredictable. It is genuinely a gamble whether or not the model will even acknowledge prompting. But if it does? It cooks well. Even seeing that this is a 32b I have a hard time believing it.

— Sellvene

[PDF omitted for brevity] This model scored higher on several components with an RP score of 67% compared to Snowdrop v0 with an RP score of 58%, especially with characterization and creativity on both text completion presets. This may be due to the effectiveness of the thinking, as responses given by this model proved to follow the plan generated by the chain of thought more closely, leading to a more in depth characterization and improved creativity.

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

I love it. Sticks to the character perfectly and some sentences hit me in right places.

— Carmenta

I like it, it's definitely meaner and hornier compared to [the other experimental models]. There are slops here and there, though for whatever reason,the responses really feel familiar. I'm starting to think if this is v0. Overall, I like it, but too horny. I prefer balance experience. Tho, if you want mean, horny model, this is the way to go.

— Raihanbook

It seems like the LLM reasoning was a bit more detailed and thoughtful somehow. Overall though if I have to give this a rating it would be 7/10.

— Sprout

Horny, 5 out of 5 swipes. I like it. While it has some inconsistency issue with formatting and length, I like the quality of responses a lot.

— OMGWTFBBQ

Excellent prose. Dialogue and narration were great!

TONS of variety when it comes to rerolls.

Much better at handling NPCs than the other models.

Reasoning can be slightly weird sometimes.

— LegendaryJerry

The responses seem to be more varied between swipes.

— br

With top nsigma 1 temp 1, the rerolls are creative, formatting holds up, reasoning's decent though falls apart rarely, and dialogue is average-to-good, following through on instructions and intro message. Feels more consistent and less creative across rerolls with regular Qwen samplers. Other testers report some user impersonation, some random Chinese/Japanese characters popping up in responses.

— 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:

• trashpanda-org/Qwen2.5-32B-Marigold-v1-ep3
• trashpanda-org/Qwen2.5-32B-Marigold-v1
• Columbidae/QwQ-32B

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-v1-ep3
    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