trashpanda-org

QwQwQwQwQwQ and Marigold met at a party and hit it off... Severian's notes : R1 at home for RP, literally. Able to handle my cards with gimmicks and subtle tricks in them. With a good reasoning starter+prompt, I'm getting consistently-structured responses that have a good amount of variation across them still while rerolling. Char/scenario portrayal is good despite my focus on writing style, lorebooks are properly referenced at times. Slop doesn't seem to be too much of an issue with thinking enabled. Some user impersonation is rarely observed. Prose is refreshing if you take advantage of what I did (writing style fixation). I know I said Marigold would be my daily driver, but this one is that now, it's that good. Context/instruct template : ChatML. Was definitely not tested with ChatML instruct and Mistral v7 template, nuh-uh. Samplers : temperature at 0.9, minp at 0.05, topa at 0.3, TFS at 0.75, repetitionpenalty at 1.03, DRY if you have access to it. (or not, see below.) 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/w812at.png Feel free to test whichever reasoning setup you're most comfortable with, but here's a recommendation from me. My prompt has a line that says: Deciding to fixate on that, my reasoning starter is: What this did for me, at least during testing is that it gave the reasoning a structure to follow across rerolls, seeking out that part of the prompt consistently. See below: But the responses were still varied, because the next few paragraphs after these delved into character details, so on and so forth. Might want to experiment and make your own thinking/reasoning starter that focuses on what you hope to get out of the responses for best results. Big thanks to the folks in the trashpanda-org discord for testing and sending over some logs! > PROS: > > In 10 swipes, had only two minor instances of speaking for {{user}}. (Can probably be fixed with a good prompt, though.) > > Creativity: 8/10 swipes provided unique text for 90% of the response, almost no cliché phrases. > > Takes personality of characters into account, sticking to it well. Even without a lorebook to support it was able to retain lore-specific terms and actually remember which meant which. > > NPCs: In 6/10 swipes NPC characters also partook in action, sticking to bits of information provided about them in opening message. Some of them even had their unique speech patterns. (Certain with a proper lorebook it would cook.) > > Unfiltered, graphic descriptions of fight scenes. Magic, physical attacks - everything was taken into account with no holding back. > > CONS: > > Some swipes were a bit OOC. Some swipes were bland, providing little to no input or any weight on the roleplay context. > > Out of all models I've tried recently, this one definitely has most potential. With proper prompting I think this beast would be genuinely one of the best models for unique scenarios. > It's one of the -maybe THE- best small thinking models right now. It sticks to character really well, slops are almost non-existent though they are still there of course, it proceeds with the story well and listens to the prompt. I LOVE R1 but I love snowdrop even more right now because answers feel more geniune and less agressive compared to R1. > Writes better than GPT 4.5. Overall, I think censorship is fucking up more unhinged bots and it's too tame for my liking. Another thing I noticed is that, it's sticking too much to being "right" to the character and too afraid to go off the rails. > I'm fainting, the character breakdown in it's thinking is similar like R1 does. Character handling looks amazing. Broo if a merge this good then, I'm looking forward to that QwQ finetune. > Negligible slop, no positivity bias which is good though. I like the model so far, R1 at home. > Overall, I think this is a real solid model. Cot is great, listens to my prompt extremely well. Number 1 for reasoning, honestly. And the way it portrays the character and persona details? Perfect. Narration, perfect. I have very little complaints about this model, ya'll cooked. > On my end, posivity bias isn't really there 🤔 Character and scenario portrayal is good. The prose too, I like it. Between this and Marigold, I feel like I can lean into snowboard (I mean Snowdrop) more. For now though, it is still Marigold. > It's pretty damn good. Better than Mullein, I think. > So far, it fucking SLAPS. I don't think it's tried to pull POV once yet. This model was merged using the TIES merge method using Qwen/Qwen2.5-32B as a base. The following models were included in the merge: trashpanda-org/Qwen2.5-32B-Marigold-v0 Qwen/QwQ-32B trashpanda-org/Qwen2.5-32B-Marigold-v0-exp The following YAML configuration was used to produce this model:

V0 note from Severian : This instruct variant is tamer and less unhinged than the base version, losing some ability to characterize NPCs but with further improved char/scenario portrayal, a tradeoff of sorts. We couldn't actually decide what to put out, because both are fun and good in their own way. Let us know what you think, we're looking forward to seeing people test it. The folks in the trashpanda and ArliAI discords for testing (In no particular order) The Allura folks for their Sugarquill 10k dataset (which I lightly cleaned for stuff like unicode quotes) fizz for her floyd-instruct, woke-identity, and benchmaxxing (lol) datasets Gryphe for their Sonnet3.5 RP and 4o WP datasets, which I heavily filtered for slop kalo's Opus-22k dataset, which was usable basically OOTB Norquinal for their OpenCAI dataset Dampfinchen for their Creative Writing Multiturn dataset The Recursal folks for their SCP wiki dataset (we also used some other private datasets of our own) > Base is more unhinged but I see more slops. Would be interesting to see if a merge can balance it out in a good way > > Instruct gives me more swipes that I like, it's less horny but it can definitely cook during actual smut > > I still like instruct more I think, but I appreciate how unhinged base model can be lol > Hard to send with one hand. What did you feed this model? > Reroll varies the response by a lot. It's giving Starcannon. > Tried the base version with my card. It's just a narrative card and the model makes the character portray right, it also mentions my persona detail often. Big thanks to the folks in the trashpanda-org discord for testing and sending over some logs! This model was merged using the TIES merge method using unsloth/Mistral-Small-24B-Instruct-2501 as a base. The following models were included in the merge: trashpanda-org/MS-24B-Mullein-v0 The following YAML configuration was used to produce this model:

- Developed by: [More Information Needed] - Funded by [optional]: [More Information Needed] - Shared by [optional]: [More Information Needed] - Model type: [More Information Needed] - Language(s) (NLP): [More Information Needed] - License: [More Information Needed] - Finetuned from model [optional]: [More Information Needed] - Repository: [More Information Needed] - Paper [optional]: [More Information Needed] - Demo [optional]: [More Information Needed] Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). - Hardware Type: [More Information Needed] - Hours used: [More Information Needed] - Cloud Provider: [More Information Needed] - Compute Region: [More Information Needed] - Carbon Emitted: [More Information Needed]

"You can take the user out of JanitorAI but you can't take the JLLM out of them." - OwenArli, 2025 hasnonname's trashpanda baby is getting a sequel. More JLLM-ish than ever, too. Severian's notes : No longer as unhinged as v0, so we're discontinuing the instruct version. Varied rerolls, good character/scenario handling, almost no user impersonation now. Huge dependence on intro message quality, but lets it follow up messages from larger models quite nicely. Currently considering it as an overall improvement over v0 as far as tester feedback is concerned. Still seeing some slop and an occasional bad reroll response, though. Let us know what you think, if it did well we'll spread it to other base models. Context/instruct template : Mistral V7 or V3 Tekken, but for some godforsaken reason, we found that this is (arguably) better with Llama 3 context/instruct. It's funny, stupid and insane, we don't know why this is the case. Trying out Llama 3 instruct/context on base MS24B told us it was coherent for it too in 4/5 responses, but not better than Mistral ones. As to why V1 seems to do better than base MS24B, we don't really know. Samplers : temperature at 0.8 - 1.25, minp at 0.05, topa at 0.2, smoothingfactor at 0.2. Some optional settings include repetitionpenalty at 1.03 or 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. Big thanks to the folks in the trashpanda-org discord for testing and sending over some logs! > It has some slop but when it cooks, the way it writes is so different. The words just go straight to my heart :UwU: > > In some rerolls, I notice it gives up using brackets for dialogues. V1 can cook but he cooks like me, sometimes good and sometimes he burns shit. > Really good. That llama preset [with V1] felt like discovering that birbs were government drones all along. > Dayum, liking it on the first gen. The fact that it narrates, something I missed from having been using Deepseek for almost a month. Really like this gen, calm and to the point. It's good for multiple people in a scene too, I don't think it's overly horny, which is good in my book. > v1 was interactive and novel, though it did say "waiting for your response a lot" -- before using llama context/instruct. It had trouble sticking to characters but after the llama preset it was managable. Spatial awareness was good and little to no impersonation. > writing style is really good but I'm not sure if it's following the bot's character really well. > > I'm having fun with v1 too btw, working really good rn. I'm so happy that it doesn't use my POV like... I could kiss this model rn > Character handling/spatial awareness is okay, I just need to reroll. I have a starter that has a pretty simple and straightforward prose, it picks up on it. On another note, the card has a faux stat system. It follows the rules accordingly. It doesn't do phrases/structure repetition like ms3 instruct atleast. It cooks, I'm liking it. > I didn't like any of the [blind test] models until I did the llama preset. I genuinely enjoyed v1, it's immersive and makes me want to continuously swipe and continue the story, not just do the same starter over and over. > I’m totally into Mullein V1’s language—it’s just so good. It’s like it has all the best parts of JLLM without most of the downsides. [...] barely ever tries to impersonate the user, which is pretty nice. It’s my default model for roleplaying now. (following up on an existing R1 chat with 40 messages)

This model is a fine-tuned version of Hasnonname/Qwen2.5-14B-Kebab-v0. It has been trained using TRL. This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. - TRL: 0.15.1 - Transformers: 4.49.0 - Pytorch: 2.5.1 - Datasets: 3.4.1 - Tokenizers: 0.21.1

This is a merge of pre-trained language models created using mergekit. This model was merged using the TIES merge method using unsloth/Mistral-Small-24B-Instruct-2501 as a base. The following models were included in the merge: trashpanda-org/Gullein trashpanda-org/Llama3-24B-Mullein-v1 The following YAML configuration was used to produce this model: