CoolFace
Modelpublic

ross-dev/SexyGPT-v3-Thinking-Female-gguf

sourceHugging Faceapache-2.0updated 26d agoView on Hugging Face
0likes541downloads
Model Card

<h1 style="text-align:center;font-size:2.5em;">SexyGPT-v3-Thinking-Female-gguf - Model Card</h1>

<p align="center"> <img src="https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female/resolve/main/model_image.png" alt="SexyGPT-v3-Thinking-Female Model Image" width="512px" height="512px" style="object-fit:cover;max-width:512px;max-height:512px;" /> </p>

<div align="center">

Model Status License Format Base Model Quant

*GGUF builds of the character-wearing model — give an LLM a persona and let it wear it.*

Website • GitHub • Email

</div>

Model Summary

GGUF quantizations of [SexyGPT-v3-Thinking-Female](https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female) — a persona-driven roleplay model built on Qwen3.8-27B (qwen3_5 architecture) that is trained not to imitate a character but to inhabit one. It opens every turn with a private <think> reasoning block (reasoning about the scene and the partner's message), then answers in-character as Monah, a consistent 21-year-old female persona.

These files are for CPU/GPU inference via llama.cpp and LM Studio. The full recipe, metrics, and training journey live on the safetensors card.

The philosophy — "wearing" a character. SFT hands the base model the wardrobe; the reasoning traces teach it to think in-character before speaking; a reward model + GRPO tailor the fit until the persona stops being a prompt it reads and becomes a skin it wears. The character is delivered entirely through the system prompt — swap the character sheet and the same weights wear a different persona.

Quick Facts

  • —Base Model: Qwen3.8-27B (qwen3_5), 27.8B params, 64 layers (48 linear-attention + 16 full)
  • —Format: GGUF (llama.cpp)
  • —Quantization: Q4KM (~4.92 bits/weight)
  • —File Size: ~16 GB
  • —Context Length: 262,144 tokens (native)
  • —Reasoning: native <think> channel with reasoning_effort (low / medium / xhigh)
  • —License: apache-2.0
  • —Created: August 2026

⚠️ Runtime Compatibility Note (read first)

The qwen3_5 gated-delta-net (hybrid linear + full attention) architecture is very new. Your llama.cpp / LM Studio build must include `qwen3_5` support to run these files. On an older build you may see the file load its tensors and then hang at generation (the linear-attention recurrent kernels are the least-mature path), or fail to recognize the architecture entirely.

  • —✅ Use an up-to-date llama.cpp / LM Studio (latest runtime with qwen3_5 / gated-delta-net support).
  • —🧪 If generation hangs or errors on your build, the [safetensors model](https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female) via Transformers / Unsloth is the reliable path today.
  • —🔁 Re-download if you ever see "tensor data is not within the file bounds / model is corrupted or incomplete" — that indicates a truncated download, not a bad model. Verify the file size and SHA256 against this repo.

Available Files

FileQuantSizeUse Case
qwen38-27b-monah-Q4_K_M.ggufQ4KM~16 GBBest size/quality balance; fits a 24 GB GPU fully offloaded

(Additional quants — Q5_K_M, Q8_0, Q6_K — can be provided on request.)

How to Use

LM Studio

  1. 1.Download the .gguf into your LM Studio models folder (or load it from disk).
  2. 2.Recommended settings:
SettingValue
Temperature0.7
Top P0.8
Top K20
Repeat Penalty1.05
Context Length4096–8192 (raise as needed; native max 262,144)
Max Tokens / response2048 (leave headroom — reasoning tokens count against this)
GPU Offloadall layers (-ngl 99 equivalent) — the Q4 fits a 24 GB card
Prompt templateChatML / Qwen (uses `<im_start> … <im_end>`)
  1. 1.System prompt (this is the character — required):
You are playing the female side of a one-on-one chat conversation with a man. Character: Monah, a playful, confident 21-year-old woman. Stay fully in character at all times, react to what he actually says, and keep one consistent voice. Reasoning effort: low.
The model thinks in a <think> block before each reply. LM Studio will show the reasoning and then the in-character message. If you want shorter thinking, keep "Reasoning effort: low" in the system prompt; for richer scenes use "medium".

llama.cpp (CLI)

bash
./llama-cli \
    -m qwen38-27b-monah-Q4_K_M.gguf \
    -ngl 99 \
    -c 4096 \
    --temp 0.7 --top-p 0.8 --top-k 20 --repeat-penalty 1.05 \
    -sys "You are playing the female side of a one-on-one chat conversation with a man. Character: Monah, a playful, confident 21-year-old woman. Stay fully in character. Reasoning effort: low." \
    -p "Hey you... I kept thinking about you today. What are you up to?"

llama.cpp (server / OpenAI-compatible)

bash
./llama-server -m qwen38-27b-monah-Q4_K_M.gguf -ngl 99 -c 8192 --host 0.0.0.0 --port 8080
# then POST to /v1/chat/completions with a system + user message, temperature 0.7, top_p 0.8

Recommended Generation Parameters

Modetemptop_ptop_kmax tokensreasoning_effort
In-character chat (default)0.70.8202048low
Deeper scenes0.60.95204096medium
More deterministic0.30.9201024low
Tip: reasoning tokens count against the response budget. Give ≥512 tokens of headroom, or the model can hit the ceiling mid-<think> and return an empty visible reply.

Training Journey (summary)

Produced entirely on local consumer hardware (RTX 3090 24 GB + RTX 3080 10 GB, WSL2):

  1. 1.SFT — QLoRA on Qwen3.8-27B over 308 female-persona ("Monah") conversations, ~77 % of assistant turns carrying <think> reasoning traces. Eval loss 0.782 (train 0.687, no overfit).
  2. 2.Reward Model — Bradley-Terry (Qwen3.5-9B, 61.4 % held-out pairwise) + a 4B twin (59.1 %) to score generations cheaply during RL.
  3. 3.GRPO — 100 steps with the 4B reward model live + structural rewards (well-formed think block, sane length, non-repetition); policy + RM co-resident across both GPUs.
  4. 4.Merge & Quantize — streaming 16-bit merge (numerically verified), then Q4KM GGUF.

Full details on the safetensors card.

Model Details

Model Information

PropertyValue
Model NameSexyGPT-v3-Thinking-Female-gguf
Base ModelQwen/Qwen3.8-27B
ArchitectureQwen3_5 (hybrid gated-delta-net + full attention)
Parameters~27.8 Billion
QuantizationQ4KM (GGUF)
Training FrameworkUnsloth + TRL + Hugging Face Transformers
Quantized withllama.cpp (convert_hf_to_gguf.py + llama-quantize)
DevelopersRoss Technologies AI Research Team
Release DateAugust 30, 2026
Model Version3.0

Model Developers

RoleNameContact
Lead DeveloperAndrei Rossdevops.ross@gmail.com
OrganizationRoss Technologies AI Research TeamIsrael

Model Repositories

  • —Model Hub (GGUF): https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female-gguf
  • —Model Hub (safetensors): https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female
  • —GitHub: https://github.com/ross-sec
  • —Developer Website: https://ross-developers.com

Model Evaluation

TaskMetricScoreNotes
SFT fitEval loss0.782no overfit (train 0.687)
Held-out repliesIn-character rate6 / 6clean, in-voice
Reward model (9B)Pairwise accuracy61.4 %gold vs sampled
Reward model (4B twin)Pairwise accuracy59.1 %margin +0.93
GRPOFormat disciplineimproved6/6 clean vs 4/6 SFT
Directional internal metrics on a small held-out set, not published benchmarks. Note: Q4KM quantization trades a small amount of fidelity for size; prefer the safetensors model for the highest quality.

Intended Use

Primary Use Cases

✅ Character-driven Conversational AI — persona chatbots that stay in voice ✅ Adult Game Development — NPC dialogue for 18+ games ✅ Interactive Storytelling — adult entertainment apps ✅ Local/Edge Inference — run a large persona model on a single consumer GPU

Out-of-Scope Use Cases

❌ Production AI systems without additional safety measures ❌ High-stakes decisions (medical, legal, financial) ❌ Autonomous / real-world decision making ❌ Misinformation ❌ Any exposure to minors — this is an 18+ model

Hardware Requirements

Use CaseRAMVRAMGPUStorage
Full GPU offload (Q4)16 GB~17 GBRTX 3090 (24 GB)16 GB
Partial offload32 GB10–16 GBRTX 3080 / similar16 GB
CPU only32 GB+–– (slow)16 GB

Ethical Considerations

THIS EXPERIMENTAL MODEL IS TUNED WITH EXPLICIT ADULT CONTENT! PLEASE DO NOT ABUSE!

  • —Single-character persona; small, low-diversity, English-only training data.
  • —Designed for playful, consensual adult roleplay — not aggression.
  • —The <think> channel makes the model's intent inspectable before it answers.

Recommendations: verify users are adults, use content filtering for public deployments, keep human oversight, and never expose the model to minors.

Privacy & Data

  • —Training data: private, proprietary dataset. No personal data. No inference-time collection.

Terms of Service

By using this model, you agree to:

  1. 1.Use it for intended (18+, lawful) purposes only
  2. 2.Not redistribute or publicly host it without permission
  3. 3.Comply with applicable laws and regulations
  4. 4.Indemnify Ross Technologies AI Research Team from liability
  5. 5.Not use it for illegal content, and never expose it to minors

Third-Party Components

  • —Base Model: Qwen3.8-27B (Alibaba Qwen License)
  • —llama.cpp: MIT • Unsloth / TRL / Transformers: Apache 2.0
  • —Hardware: CUDA (NVIDIA License)

Maintenance & Support

  • —Current Version: 3.0 • Release Date: August 30, 2026 • Status: Active, Maintained

Primary Contact: devops.ross@gmail.com Organization: Ross Technologies AI Research Team

Reporting issues: email devops.ross@gmail.com or comment on the model card. Best-effort response.

Citation & Attribution

bibtex
@model{sexygpt_v3_thinking_female_gguf_2026,
  title={SexyGPT-v3-Thinking-Female (GGUF): A Character-Wearing 27B Roleplay Model},
  author={Ross, Andrei},
  organization={Ross Technologies AI Research Team},
  year={2026},
  howpublished={\url{https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female-gguf}}
}

Acknowledgments

  • —Alibaba Qwen Team — Qwen3.8 base model and thinking capabilities
  • —Unsloth — QLoRA/GRPO training stack on consumer GPUs
  • —ggml / llama.cpp — GGUF format and quantization tooling
  • —Hugging Face — hub and transformers/TRL libraries

Contact Information

📧 Email: devops.ross@gmail.com 🌐 Website: https://ross-developers.com 💻 GitHub: https://github.com/ross-sec

Developer: Andrei Ross (devops.ross@gmail.com) Organization: Ross Technologies AI Research Team

Legal Disclaimer

This model is provided "AS IS" without warranty of any kind. Ross Technologies AI Research Team makes no representations about the model's suitability for any particular purpose. Users are solely responsible for determining the appropriateness of use and assume all risks associated with deployment.

Model Card Version: 3.0 Last Updated: August 30, 2026 Created by: Ross Technologies AI Research Team

For the most current version and updates, visit: https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female-gguf