Forgotten-Safeword

Adults only (18+). These models are unaligned and produce explicit sexual content, including taboo themes, with no refusals. You are responsible for how you use them and for what they generate.

Forgotten-Safeword is a family of roleplay models by ReadyArt (sleepdeprived3), made for unrestricted erotic roleplay (ERP) and dark creative writing. Each size is an existing roleplay model further trained on ReadyArt’s Safeword dataset, which covers 23 kinds of taboo content. This page collects GGUF builds of every size for Ollama, with each size and quantization as its own tag.

The sizes

Available now: 12B, 36B and 70B. The 8B, 22B and 24B are coming soon.

How they differ

• Mistral or Llama. The 8B to 36B are built on Mistral models, the 70B on Meta’s Llama 3.3. They use different chat formats; Ollama applies the right one automatically, but it matters if you use your own prompt template or a frontend preset (pick a Mistral preset for 8B–36B and a Llama 3 preset for 70B).
• Different roleplay bases. Each size starts from a different roleplay finetune (TheDrummer’s Ministrations, UnslopNemo, UnslopSmall, Cydonia and Skyfall; SicariusSicariiStuff’s Negative_LLAMA), so they don’t just differ in size but also in writing style.
• Bigger is smarter, slower and hungrier. Larger models follow complex scenes, characters and instructions better and keep track of details longer, but need far more memory. The 22B and 24B are different models of similar size: the 24B is built on the newer Mistral Small. The 36B is an enlarged 24B.
• Versions. The 70B is the latest release (v5.0); the 36B is v4.1 and the others v4.0. These are the last versions ReadyArt released for each size.

Quantizations

Every size is available in several quantizations. Quantization stores the model’s weights with fewer bits, which makes it smaller and faster at the cost of some quality. The tag says which one you get: <size>-<quantization>, for example 24b-q4_K_M.

A larger model at a lower quantization often beats a smaller model at a higher one: a 70B at IQ3_M usually writes better than a 36B at Q8_0. Low quantizations hurt small models more, which is why the 8B and 12B only come in Q4_K_M and up.

All versions except Q8_0 are made with an importance matrix (imatrix), which keeps the most important weights more precise and makes the smaller versions noticeably better.

Sizes

Plain <size> tags (:12b, :36b, :70b …) are the Q8_0 version. latest will be the 8B Q8_0 once that is published. Sizes of versions that aren’t published yet are estimates.

A dash means that size isn’t offered in that quantization: small models lose too much at low bit rates, and only the 70B is big enough for IQ2_M to make sense.

You need roughly the file size plus 1–3 GB of combined VRAM and RAM at the default 8K context (more for bigger models and longer contexts). The more of the model fits in your GPU’s VRAM, the faster it runs: a model that fits entirely in VRAM is many times faster than one split between GPU and CPU.

Usage

ollama run leeplenty/forgotten-safeword:70b-q4_K_M
ollama run leeplenty/forgotten-safeword:36b-q4_K_M
ollama run leeplenty/forgotten-safeword:12b

Defaults set in these models:

• Sampling settings recommended for each size: for the Mistral-based sizes the samplers from the SillyTavern preset on the model card (12B: temperature 0.7, 36B: temperature 1.0, both with top_k 40, min_p 0.05); for the 70B its generation config (temperature 0.6, top_p 0.9).
• num_ctx 8192. The models support much longer contexts; raise it with /set parameter num_ctx 16384 or in your client, at the cost of more memory.

For SillyTavern, the model cards recommend sleepdeprived3’s presets: Mistral-V3-Tekken-E for the 12B, Mistral-V7-Tekken-E for the 36B and Llama-3.3-T4 for the 70B.

How these files were made

Converted and quantized from the original BF16 safetensors with llama.cpp, using the importance matrices published by mradermacher for each model.

Credits

• sleepdeprived3 / ReadyArt: Forgotten-Safeword author
• TheDrummer: Ministrations, UnslopNemo, UnslopSmall, Cydonia and Skyfall, the bases of the 8B–36B (Ko-fi)
• SicariusSicariiStuff: Negative_LLAMA, the base of the 70B (Ko-fi)
• mradermacher: importance matrices
• Mistral AI and Meta: the underlying base models
• Community: Beaver AI Discord

All credit for the models goes to their creators; this page only provides Ollama builds.

License

Each size follows the license on its original model card. The 12B and 36B are Apache 2.0. The 70B is built with Llama and is subject to the Llama 3.3 Community License and Meta’s Acceptable Use Policy.