allura-org

main { --creep-bg: #0f0f0f; --blood-rust: #5e2e28; --faded-white: #e8e8e8; background: var(--creep-bg); border-radius: 10px; } main, details { display: flex; flex-direction: column; align-items: center; padding: 15px; overflow-x: scroll; scrollbar-width: none; } .warning-box { background: #2d1a1a; border: 1px solid var(--blood-rust); padding: 15px; margin: 20px 0; position: relative; overflow: hidden; } .warning-box::before { content: ''; position: absolute; top: 0; left: -10%; width: 120%; height: 100%; background: linear-gradient(90deg, transparent 0%, #ff000020 50%, transparent 100%); animation: scan 4s infinite linear; } .content-block { background: #1a1a1a; border: 1px solid #333; border-radius: 4px; padding: 20px; margin: 15px 0; box-shadow: 0 2px 8px rgba(0,0,0,0.3); position: relative; } code { background: #000; color: #f8f8f8; padding: 4px 6px; border-radius: 3px; font-family: monospace; border: 1px solid #333; } .spoiler-text { background: linear-gradient(45deg, #ff3b3b, #ff6b6b); -webkit-background-clip: text; color: transparent; text-shadow: 0 0 12px rgba(255,59,59,0.5); animation: glitch 1s infinite steps(2); } img { max-width: min(90vw, 400px); border-radius: 6px; filter: brightness(0.95) contrast(1.1); } @media (max-width: 768px) { main { padding: 10px; } .content-block { padding: 15px; margin: 10px 0; } details[open] { padding-bottom: 20px; } } @keyframes glitch { 0% { transform: translateX(0); } 25% { transform: translateX(-1px); } 50% { transform: translateX(1px); } 75% { transform: translateX(-1px); } 100% { transform: translateX(0); } } @keyframes scan { 0% { transform: translateX(-20%); } 100% { transform: translateX(120%); } } Us here at Playtime Co. are excited to welcome you to our QA staff! As a signatory of the Employee Confidentiality Agreement, you are required to keep all information regarding the company and our research confidential. A̶͇͊r̵͎̭͗͛e̶̯̘̋ ̵̤͖̄̕y̶̩̎o̵̡̿u̴̙̅͂ ̵̖̋r̷̮͖̓ě̵̪̹̈́a̸̠̅d̵͖̖̔̋ẙ̷͙̃ ̷̮͖̋̈́t̴͉̅́ȍ̴̹̞ ̵̖̜̈́p̵͕̑̀ľ̵̥̓ȁ̷͔̩̇y̷̞͊̿ ̸͍̺̀͝w̶͙͈̑i̷̡͗̾t̸͎͒̐ḩ̸̳̓ ̴̮̺̇ȗ̴̢͈̉ş̷͖̔̒?̵̝̺́ This model was designed for exceptional skill in roleplaying, both with adults and children, for our patent pending Playtime Playground. It seems to have succeeded. Native-quality English Mediocre Chinese (further research needed) <s>[SYSTEMPROMPT]What to roleplay as[/SYSTEMPROMPT][INST]User: xxx[/INST]ASSISTANT: yyy</s> Fun to use, nice swipe variation, gives me lots to RP off of. Rarely, it'll start to loop, but a quick swipe fixes no problem. It passes my silly logic tests (read: me trolling random characters) Haven't seen any slop yet Writes short and snappy replies ...yet not too short, like Mahou, and can write longer responses if the context warrants it Follows card formatting instructions If this holds up to 16K it will be constantly in the hopper alongside Mag-Mell for me. I'm biased towards shorter responses with smarts. :) tantalizing writing, leagues better then whatever is available online ⚠️ WARNING While Experiment 12B has shown exceptional performance, researchers must maintain safety protocols. Standard containment procedures apply. Playtime Co., Poppy Playtime, Huggy Wuggy, and all related properties are trademarks of Mob Entertainment LLC. Not affiliated with or endorsed by Mob Entertainment. Obviously, children should not be using it. The mention up there was a part of the bit. This is all a bit!!!

This is a 50/50 merge of two separate trains: - ToastyPigeon/g3-12b-rp-system-v0.1 - ~13.5M tokens of instruct-based training related to RP (2:1 human to synthetic) and examples using a system prompt. - ToastyPigeon/g3-12b-storyteller-v0.2-textonly - ~20M tokens of completion training on long-form creative writing; 1.6M synthetic from R1, the rest human-created Update: Vision has returned to this model, rejoice. Uses Gemma2/3 instruct, but has been trained to recognize an optional system role.

But! Unlike the last couple! This one actually doesn't suck at RP or adventure! Hooray! Usage notes: Chat template is ChatML, since this was trained using Muse-12B as the base model. Works with Temperature 1 / min-p 0.05, up through about Temperature 1.3 / min-p 0.02 if you like it hotter, probably works with other settings too. Training notes: I started with Muse-12B because I like, in general, how this model behaves for creative uses. It's been my daily driver for a while, but I felt like the prose could stand to be a little more variable. Stage 1: - Data: A bunch of books (chosen for their prose/writing style). About 28M tokens per epoch. - r32/a32 QLoRA at 32k context, applied only to the QKV tensors. LR: 1e-5, 2 epochs. Stage 2: - Trained on top of Stage 1. - Data: RP data, about 4M tokens. - r32/a32 QLoRA at 16k context, applied to `oproj` and `downproj` only. LR: 5e-6, 1 epoch. Stage 3: - Trained on top of Stage 2. - Data: Instruct data (1000 random samples from koto-instruct-sft), about 1.2M tokens. - r32/a32 QLoRA at 4k context, applied to all linear modules. LR: 2e-6, 1 epoch. (Big thanks to Gryphe and LatitudeGames for making Muse-12B. Great model, knocked it out of the park fr)

Static GGUF quants of TQ2.5-14B-Sugarquill-v1, made with llama.cpp version `b4052`

A roleplay-focused LoRA finetune of Qwen 2.5 32b Instruct. Methodology and hyperparams inspired by SorcererLM and Slush. Yet another model in the Ink series, following in the footsteps of the Nemo one Testimonials > whatever I tested was crack [...] It's got some refreshingly good prose, that's for sure > The NTR is fantastic with this tune, lots of good gooning to be had. [...] Description and scene setting prose flows smoothly in comparison to larger models. > This 32B handles complicated scenarios well, compared to a lot of 70Bs I've tried. Characters are portrayed accurately. > From the very limited testing I did, I quite like this. [...] I really like the way it writes. > Granted, I'm completely shitfaced right now, but I'm pretty sure it's good. > [This model portrays] my character card almost exactly the way that I write them. It's a bit of a dream to get that with many of the current LLM. Dataset The worst mix of data you've ever seen. Like, seriously, you do not want to see the things that went into this model. It's bad. "this is like washing down an adderall with a bottle of methylated rotgut" - inflatebot Recommended Settings Chat template: ChatML Recommended samplers (not the be-all-end-all, try some on your own!): - Temp 0.85 / Top P 0.8 / Top A 0.3 / Rep Pen 1.03 - Your samplers can go here! :3 Hyperparams General - Epochs = 1 - LR = 6e-5 - LR Scheduler = Cosine - Optimizer = Paged AdamW 8bit - Effective batch size = 16 LoRA - Rank = 16 - Alpha = 32 - Dropout = 0.25 (Inspiration: Slush) Credits Humongous thanks to the people who created the data. I would credit you all, but that would be cheating ;) Big thanks to all Allura members, for testing and emotional support ilya /platonic especially to inflatebot who made the model card's image :3 Another big thanks to all the members of the ArliAI Discord server for testing! All of the people featured in the testimonials are from there :3

Koto-Small-7B-PT is a version of MiMo-7B-Base trained on almost a billion tokens of creative writing data. Please check out Aurore-Reveil/Koto-Small-7B-IT, it's the official RP and instruct tune! This model is not intended for use outside of raw text completion settings, such as cowriting. Instruct will not work. Multi-turn roleplay will not work. It was trained at 32k, but as not all samples were this long, we expect that in the best case you can get ~16k effective context. We found that 1.25 temperature and 0.05 minp worked best, but YMMV! Some of the data used to train this model includes: - Most of The Anarchist Library, a repository for anarchist manifestos and writing (see allura-org/the-anarchist-library) - A random sample of public domain books from Project Gutenberg - Furry (anthro and feral) storytelling and smut - A small subset of known high-quality books and story data Acknowledgements - thank you to [unk] for drawing the art used in the model card! - thank you very much to mango/deltavector for providing the compute used to train this model - thanks to curse for testing, ideas - thanks to toasty for some data, ideas - thanks to everyone else in allura for moral support Training Notes This model was trained over the course of ~18 hours on an A100 node. We used 8-bit AdamW and the Cosine LR scheduler, as well as both gradient clipping and weight decay for regularization. Before training, we converted the original model to the Qwen 2 architecture by removing the MTP weights and custom modelling code, and slightly modifying the `config.json`. This allowed us to use CCE and Liger which let the train go much faster than it would have otherwise. We decided to keep the final model in the converted Qwen 2 format, as it is more supported by community software such as EXL2, EXL3, Aphrodite, etc, as well as the original architecture's MTP weights likely being much less effective after finetuning without them. Finetuning Notes This model has had ChatML tokens already added by Xiaomi. Please use this format when finetuning to ensure compatibility with the rest of the ecosystem.

A merge of Sugarquill and Sunfall. I wanted to combine Sugarquill's more novel-like writing style with something that would improve it's RP perfomance and make it more steerable, w/o adding superfluous synthetic writing patterns. I quite like Crestfall's Sunfall models and I felt like Gemma version of Sunfall will steer the model in this direction when merged in. To keep more of Gemma-2-9B-it-SPPO-iter3's smarts, I've decided to apply Sunfall LoRA on top of it, instead of using the published Sunfall model. I'm generally pleased with the result, this model has nice, fresh writing style, good charcard adherence and good system prompt following. It still should work well for raw completion storywriting, as it's a trained feature in both merged models. Thanks to Prodeus, Inflatebot and ShotMisser for testing and giving feedback. Model responds to Gemma instruct formatting, exactly like it's base model. The following YAML configuration was used to produce this model:

there was originally going to be a better logo but i couldnt get any image model working. so this is what you all deserve Lune Mamba 3B GRPOIF is a Claude-OSS series model based on Granite 4.0 H(ybrid) Micro. Claude-OSS is a (non-affiliated with Anthropic!) attempt to replicate the style of Anthropic's Claude model on top of open source bases. Benchmarks | Granite 4.0 H Micro | Lune Mamba 3B | Lune Mamba 3B GRPOIF -|-|-|- MMLU|63.7860|64.2338|64.3443 IFEval|80.2218|75.0462|77.4492 IFEval numbers calculated from prompt loose accuracy Artifacts - SFT checkpoint: allura-forge/claumba-micro-sft - KTO checkpoint: allura-org/Lune-Mamba-3B-v1 - GRPO (on IFeval) checkpoint: You are here!

A roleplay-focused LoRA finetune of Llama 3.1 8B Instruct. Methodology and hyperparams inspired by SorcererLM and Slush. Yet another model in the Ink series, following in the footsteps of the rest of them Dataset The worst mix of data you've ever seen. Like, seriously, you do not want to see the things that went into this model. It's bad. "this is like washing down an adderall with a bottle of methylated rotgut" - inflatebot Update: I have sent the (public datasets in the) data mix publicly already so here's that Recommended Settings Chat template: Llama 3.1 Recommended samplers (not the be-all-end-all, try some on your own!): - Temp 1.03 / Top A 0.3 / TFS 0.75 / Rep Pen 1.03 - Your samplers can go here! :3 Hyperparams General - Epochs = 2 - LR = 6e-5 - LR Scheduler = Cosine - Optimizer = Paged AdamW 8bit - Effective batch size = 16 LoRA - Rank = 16 - Alpha = 32 - Dropout = 0.25 (Inspiration: Slush) Credits Humongous thanks to the people who created and curated the original data Big thanks to all Allura members, for testing and emotional support ilya /platonic especially to inflatebot who made the model card's image :3

there was originally going to be a better logo but i couldnt get any image model working. so this is what you all deserve Lune Mamba 3B is a Claude-OSS series model based on Granite 4.0 H(ybrid) Micro. Claude-OSS is a (non-affiliated with Anthropic!) attempt to replicate the style of Anthropic's Claude model on top of open source bases. Benchmarks | Granite 4.0 H Micro | Lune Mamba 3B | Lune Mamba 3B GRPOIF -|-|-|- MMLU|63.7860|64.2338|64.3443 IFEval|80.2218|75.0462|77.4492 IFEval numbers calculated from prompt loose accuracy Artifacts - SFT checkpoint: allura-forge/claumba-micro-sft - KTO checkpoint: You are here! - GRPO (on IFeval) checkpoint: allura-org/Lune-Mamba-3B-v1-GRPOIF

There's a wisp of dust in the air. It feels like its from a bygone era, but you don't know where from. It lands on your tongue. It tastes nice. Remnant is a series of finetuned LLMs focused on SFW and NSFW roleplaying and conversation. Recommended Settings Chat template: ChatML. Apparently Llama 3 format works too, though? Ymmv :3 Samplers: - `0.8` temp - `0.1` minp - `0.5` presence penalty Credits Humongous thanks to Allura, ilya ](https://github.com/axolotl-ai-cloud/axolotl) See axolotl config

She looked into His Spine, into His Heart; and she saw there the shade of His soul. Intended as a direct upgrade to Pentiment, Q3-30B-A3B-Designant is a roleplaying model finetuned from Qwen3-30B-A3B-Base. During testing, Designant punched well above its weight class in terms of active parameters, demonstrating the potential for well-made lightweight Mixture of Experts models in the roleplay scene. While one tester observed looping behavior, repetition in general was minimal. Warning: Quantization seems very janky with Qwen 3 MoE models. We recommend using full bf16 weights and vLLM, if possible. GGUF: - imatrix GGUFs by Bartowski - Linear GGUFs by mradermacher Some users report even more issues with low-bit GGUF quants for Qwen3 MoE models. We'd recommend trying both imatrix and linear, as well as q5+ for proper quality. - Format is plain-old ChatML (please note that, unlike regular Qwen 3, you do not need to prefill empty think tags for it not to reason -- see below). - Settings used by testers varied, but Fizz and inflatebot used the same settings and system prompt recommended for GLM4-32B-Neon-v2. - The official instruction following version of Qwen3-30B-A3B was not part of the merge. Instruction-following is trained in post-hoc, and "thinking" traces were not included. As a result of this, "thinking" will likely not function as intended. - As with any Q3-30B-A3B, Designant performs very adequately with few or zero layers offloaded to GPU. When using the ikllama.cpp server, a 7950X CPU with 32GB of DDR5 RAM can run a Q4KM quant of this architecture at ~15 tokens/sec with no GPU involved at all. 1. The base model first went through a supervised finetune on a corpus of instruction following data, roleplay conversations, and human writing based on the Ink/Bigger Body/Remnant lineage. 2. It was then slightly merged with Pantheon-Proto-RP-1.8, to improve stability. 3. Finally, a KTO reinforcement learning phase steered the model away from the very purple prose the initial merge had, and improved its logical+spatial reasoning and sense of overall "intelligence". - Toaster, OMGWTFBBQ, The Trashpanda Testing Crew - Testing - inflatebot - Model Card, Testing, Merging Consultation - Gryphe, Alibaba - Making the original models as well as the ones used in the merge Bot would like to thank the Allura community on Discord, especially Curse, Vagabond, Artus and Mawnipulator, for their companionship and moral support. You all mean the world to us.

There's a wisp of dust in the air. It feels like its from a bygone era, but you don't know where from. It lands on your tongue. It tastes nice. Remnant is a series of finetuned LLMs focused on SFW and NSFW roleplaying and conversation. Quants GGUF (IF YOU ARE USING KOBOLD.CPP, PLEASE USE `--nobostoken`.): - bartowski's imatrix quants Recommended Settings Chat template: GLM4 Samplers: - `1.0` temp - `0.1` minp Credits Humongous thanks to Allura, ilya ](https://github.com/axolotl-ai-cloud/axolotl) See axolotl config

get it? because kintsugi sounds like kitsune? hahaha- Q3-8B-Kintsugi is a roleplaying model finetuned from Qwen3-8B-Base. During testing, Kintsugi punched well above its weight class in terms of parameters, especially for 1-on-1 roleplaying and general storywriting. MLX: - 8, 6, and 4bpw MLX-formrt quants by soundTeam - Format is plain-old ChatML (please note that, unlike regular Qwen 3, you do not need to prefill empty think tags for it not to reason -- see below). - Settings used by testers varied, but we generally stayed around 0.9 temperature and 0.1 min p. Do not use repetition penalties (DRY included). They break it. - Any system prompt can likely be used, but I used the Shingame system prompt (link will be added later i promise) - The official instruction following version of Qwen3-8B was not used as a base. Instruction-following is trained in post-hoc, and "thinking" traces were not included. As a result of this, "thinking" will not function. 1. The base model first went through a supervised finetune on a corpus of instruction following data, roleplay conversations, and human writing based on the Ink/Bigger Body/Remnant lineage. 2. Finally, a KTO reinforcement learning phase steered the model away from the very purple prose the initial merge had, and improved its logical+spatial reasoning and sense of overall "intelligence". Both stages here are very similar to Q3-30B-A3B-Designant, which went through a very similar process with the same data. - Toaster, Mango, Bot, probably others I forgot ;-; - Testing - inflatebot - original Designant model card that this one was yoinked from - Axolotl, Unsloth, Huggingface - Making the frameworks used to train this model (Axolotl was used for the SFT process, and Unsloth+TRL was used for the KTO process) - All quanters, inside and outside the org, specifically Artus, Lyra, and soundTeam/Heni We would like to thank the Allura community on Discord, especially Curse, Heni, Artus and Mawnipulator, for their companionship and moral support. You all mean the world to us <3

This model is a preview, unfinished, and still in development. It is not representative of any final product and has only been remotely published to prove that I am doing something productive with my life Koto-Large-106B-Preview is a version of Ling-Flash-Base-2.0 trained on almost a billion tokens of creative writing data. Thanks to lium.io for the compute! Temp 1.1, minp 0.01, rep pen 1.02, freq pen -0.04 Datasets Some of the data used to train this model includes: - Most of The Anarchist Library, a repository for anarchist manifestos and writing (see allura-org/the-anarchist-library) - A random sample of public domain books from Project Gutenberg - Furry (anthro and feral) storytelling and smut - A small subset of known high-quality books and story data Acknowledgements - thanks again to fish and co from lium for compute - thanks to curse for testing, ideas - thanks to toasty for some data, ideas - thanks to everyone else in allura for moral support Yeah idek what went wrong here, I'll be for real. It's... oddly really stupid, though it reportedly outputs good prose sometimes. It's similar to GLM-4 32B's base in that nature. Took ~18hrs on 8xH200. intervitens had already converted the model to use a faster MoE layer for training and I further patched it to use CCE to bring down memory by a little bit. TorchAO's 8bit Adamw was used for optimization. FSDP was utilized for model sharding (it's a lot more stable and supported than Deepspeed, ime) - MoE training is busted (possible, but the 0.25 epoch checkpoint looked oddly promising, so I don't think that's it) - The model did not get trained on enough data - The model was not trained aggressively enough I have a feeling that higher LR and/or doing more epochs on the data would result in a much better end result.

DO NOT USE THIS MODEL. DO NOT QUANT THIS MODEL. THE RELEASE VERSION IS PROBABLY BETTER initial version of koto trained on an earlier version of the dataset has a slightly different flavor than the release model. works best at ~1.15 temp and 0.01-0.02 minp

Koto-22B-PT is a depth-upscaled version of Mistral-Nemo-Base-2407, healed and trained on almost a billion tokens of creative writing data. This model is not intended for use outside of raw text completion settings, such as cowriting. Instruct will not work. Multi-turn roleplay will not work. It was trained at 32k, but as not all samples were this long, we expect that in the best case you can get ~16k effective context. We found that 1.5-1.55 temperature and 0.05-0.1 minp worked best, but YMMV! Some of the data used to train this model includes: - Most of The Anarchist Library, a repository for anarchist manifestos and writing (see allura-org/the-anarchist-library) - A random sample of public domain books from Project Gutenberg - Furry (anthro and feral) storytelling and smut - A small subset of known high-quality books and story data - thank you to @takeshimaxfj on twitter for drawing the art used in the model card! - thank you very much to mango/deltavector for providing the compute used to train this model - thanks to curse for testing, ideas - thanks to toasty for some data, ideas - thanks to everyone else in allura for moral support This model was trained over the course of ~14 hours on an 8xB200 node. We used 8-bit AdamW and the REX LR scheduler, as well as both gradient clipping and weight decay for regularization. There was a very odd loss spike ~60% of the way through training, but it recovered and the model seems fine? So? Eh? If it works it works :3 This model has had ChatML tokens already added if you prefer to tune using that chat format. Please do not readd them to maintain the vocab size for (possible) usage on places like Featherless

A roleplay-focused LoRA finetune of Teleut 7b. Methodology and hyperparams inspired by SorcererLM and Slush. Dataset The worst mix of data you've ever seen. Like, seriously, you do not want to see the things that went into this model. It's bad. Recommended Settings Chat template: ChatML Recommended samplers (not the be-all-end-all, try some on your own!): - Temp 1.03 / TopK 200 / MinP 0.05 / TopA 0.2 - Temp 1.03 / TFS 0.75 / TopA 0.3 Quants - Static GGUFs (thanks auri!) - Imatrix GGUFs (thanks bart!) Hyperparams General - Epochs = 2 - LR = 6e-5 - LR Scheduler = Cosine - Optimizer = Paged AdamW 8bit - Effective batch size = 12 LoRA - Rank = 16 - Alpha = 32 - Dropout = 0.25 (Inspiration: Slush) Credits Humongous thanks to the people who created the data. I would credit you all, but that would be cheating ;) Big thanks to all Allura members, especially Toasty, for testing and emotional support ilya /platonic NO thanks to Infermatic. They suck at hosting models

not too sure how I feel about this one, but yolo! :3 Triple stage RP/general tune of Qwen3-30B-A3b Base (finetune, merged for stablization, aligned) Format use chatml. thinking may or may not work, ymmv!. Thankses special thanks to alibaba for training the base model and regular instruct model, as well as Gryphe for training the pantheon model also used in the merging step. special thanks to artus for making the exllama quants. special thanks to allura for being cute :3 Postmortem never merge with qwen 3 instruct. it's not worth it. it will destroy your model and make it just qwen 3 instruct again with all its issues.