ross-dev/SexyGPT-v3-Thinking-Female

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

*Give a language model a character — and teach it to wear the character.*

Website • GitHub • Email

</div>

Model Summary

SexyGPT-v3-Thinking-Female is a persona-driven conversational model that does not merely imitate a character — it inhabits one. Built on Qwen3.8-27B (the qwen3_5 gated-delta-net architecture), it was taught, through a full three-stage pipeline, to think as its character before it ever speaks as her.

The character is Monah — a 21-year-old woman written to hold a single, consistent voice across intimate one-on-one chat. Every assistant turn opens with a private <think> reasoning block where the model reasons about the scene, the character's constraints, and the partner's last message, then closes the block and answers in-character. The result is a model that keeps the costume on: it stays in voice, follows the scene's rules, and responds to what was actually said rather than reciting canned lines.

The philosophy — "wearing" a character. A base LLM knows how to talk. Our pipeline teaches it who it is. Supervised fine-tuning hands it the wardrobe. The reasoning traces teach it to think in-character before speaking. A reward model plus GRPO then tighten the fit — rewarding replies that stay in voice and on-scene — until the character stops being a prompt the model reads and becomes a skin it wears.

Quick Facts

• —Base Model: Qwen3.8-27B (qwen3_5 architecture)
• —Model Size: ~52 GB (bf16 merged weights) / 27.8B parameters
• —Architecture: Qwen3.5 hybrid transformer (64 layers, gated-delta-net linear attention + periodic full attention)
• —Fine-tuning Method: QLoRA SFT → Reward Model → GRPO → 16-bit merge
• —Reasoning: Native <think> blocks with reasoning_effort control (low / medium / xhigh)
• —Context Length: 262,144 tokens
• —License: apache-2.0
• —Created: August 2026

Model Details

SYSTEM PROMPT

The character is delivered entirely through the system prompt — this is the "costume." A minimal form:

{"role": "system", "content": "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 was trained across several scene "modes" (e.g. teasing/foreplay, in-the-moment, phone, afterglow, and a work-assistant mode), each carrying its own pacing rules in the system prompt. Swap the character sheet in the system prompt and the same weights will wear a different character — the pipeline teaches the skill of inhabiting a persona, not just one fixed persona.

Model Information

Model Developers

Model Repositories

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

Model Architecture

Architecture Details

Qwen3_5ForConditionalGeneration (text path)
├─ Vocabulary Size: 248,320 tokens
├─ Hidden Size: 5,120 dimensions
├─ Number of Layers: 64 transformer blocks
│    ├─ 48 × linear attention (gated delta-net)
│    └─ 16 × full attention  (every 4th layer)
├─ Head Dimension: 256
├─ Attention Output Gate: enabled
├─ Intermediate Size (FFN): 17,408 dimensions
├─ Max Position Embeddings: 262,144
├─ Activation: SiLU (Swish)
├─ Normalization: RMSNorm
├─ Chat Template: ChatML + <think> reasoning channel
└─ Precision: BFloat16

The qwen3_5 hybrid design is why this model is efficient at long context (most layers use linear-attention delta-net rather than quadratic full attention) and why it opens every assistant turn with a reasoning block by default.

How to Use

Important: on Ampere GPUs (RTX 30xx) set UNSLOTH_COMPILE_DISABLE=1 and DISABLE_LLVM_OPT=1 in the environment before running — the qwen3_5 decode path otherwise hits a Triton/LLVM crash. The 27B loads in 4-bit across a single 24 GB card, or split over two GPUs.

Quick Start (Hugging Face / Unsloth)

import os
os.environ["UNSLOTH_COMPILE_DISABLE"] = "1"
os.environ["DISABLE_LLVM_OPT"] = "1"

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    "ross-dev/SexyGPT-v3-Thinking-Female",
    max_seq_length=4096,
    load_in_4bit=True,
)

messages = [
    {"role": "system", "content": "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."},
    {"role": "user", "content": "Hey you... I kept thinking about you at work today. What are you up to?"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    reasoning_effort="low",   # low | medium | xhigh
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, top_p=0.8, top_k=20, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

The reply arrives as {reasoning}\n</think>\n\n{visible answer} — split on the last </think> to separate the private thinking from the in-character message.

Using with GGUF (llama.cpp / LM Studio)

See the dedicated GGUF repository: https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female-gguf (please read its runtime note — qwen3_5 gated-delta-net requires up-to-date llama.cpp support).

Generation Parameters (Recommended)

For In-Character Roleplay (default)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.7,
    top_p=0.8,
    top_k=20,
    do_sample=True,
)  # reasoning_effort="low" in apply_chat_template

For Deeper Reasoning

outputs = model.generate(
    **inputs,
    max_new_tokens=4096,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
    do_sample=True,
)  # reasoning_effort="medium" or "xhigh"

Tip: reasoning tokens count against max_new_tokens. Give the model ≥512 tokens of headroom or it can hit the ceiling mid-<think> and return an empty visible reply. Use reasoning_effort="low" for snappy chat.

Training Details — The Journey

This model was produced entirely on local consumer hardware (RTX 3090 24 GB + RTX 3080 10 GB, WSL2), no cloud, through a deliberate "dress the character, then tailor the fit" pipeline:

Stage 1 — Supervised Fine-Tuning (the wardrobe)

• —Base: unsloth/Qwen3.8-27B in 4-bit, QLoRA (rank 8, all linear modules), batch 1, gradient checkpointing, split across both GPUs (sequential model-parallel).
• —Data: a curated private roleplay dataset — the female/Monah side, 308 deduplicated multi-turn conversations across five scene modes, of which ~77 % of assistant turns carry chain-of-thought reasoning traces that render into the model's <think> channel.
• —Result: train loss 0.687, eval loss 0.782 with no train/eval gap — the character fit cleanly without overfitting.

Stage 2 — Reward Model (learning what "in-character" means)

• —Generated multiple candidate replies per context from the SFT model, paired each against the human-written gold reply, and trained a Bradley-Terry reward model (Qwen3.5-9B + scalar head) to prefer in-character, on-scene replies.
• —Held-out pairwise accuracy: 61.4 % (mean margin +0.45). A distilled Qwen3.5-4B twin (59.1 %) was trained to score cheaply enough to sit beside the 27B policy during RL.

Stage 3 — GRPO (tailoring the fit)

• —Group Relative Policy Optimization (TRL) continued the SFT policy for 100 steps, with the 4B reward model scoring generations live plus structural rewards for a well-formed <think> block, sane reply length, and non-repetition.
• —The 27B policy and the 4B reward model ran co-resident across the two GPUs (weights split 19 GB / 5 GB, generation on the Unsloth path). GRPO tightened format discipline and voice consistency over the SFT baseline.

Stage 4 — Merge & Quantize (packaging)

• —The GRPO adapter was merged into 16-bit weights via a streaming, memory-bounded merge (peak ~20 GB RAM — no 56 GB spike), then numerically verified (LoRA delta exact, norms bit-identical) and quantized to Q4_K_M GGUF.

Training Dataset

• —Source: private, proprietary roleplay dataset (dataset_last.jsonl)
• —Split: conversation-level train/eval (no conversation straddles the split)
• —Fields: messages (system + strict user/assistant turns), per-turn reasoning_data (chain-of-thought), type (scene mode)
• —Format: ChatML + <think> reasoning channel (native Qwen3.5 template)

Model Evaluation

Evaluation Methodology

Model evaluated on:

• —Character Consistency: does it hold one voice across a conversation?
• —Scene-Rule Compliance: does it follow the per-mode pacing rules in the system prompt?
• —Responsiveness: does it react to the partner's actual last message?
• —Reasoning Quality: are the <think> traces coherent and in-character before the reply?
• —Format Discipline: well-formed think/answer channel, no empty or truncated replies.

Results (measured on this pipeline)

Metrics reflect our internal held-out evaluation on a small (44-pair / 6-context) set — treat them as directional signal, not published benchmarks.

Limitations & Known Issues

Model Limitations:

• —Trained on a small, low-diversity persona set (one character, ~308 conversations) — it is a specialist, not a generalist.
• —Reasoning traces are tuned for roleplay planning, not factual/mathematical reasoning.
• —English only.
• —Reply quality is highest at reasoning_effort="low"; higher effort can spend the token budget inside <think>.

Runtime note (GGUF):

• —The qwen3_5 gated-delta-net architecture is very new. As of the tooling used at release, llama.cpp / LM Studio may not yet run the GGUF (generation can hang on the linear-attention recurrent kernels). The safetensors model via Transformers/Unsloth is the reliable path today. See the GGUF card for details.

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 ✅ Research: persona conditioning, reasoning-augmented roleplay, RLHF/GRPO on consumer GPUs

Out-of-Scope Use Cases

❌ Production AI Systems: without additional safety measures ❌ High-Stakes Decisions: medical, legal, financial advice ❌ Autonomous Systems: real-world decision making ❌ Misinformation: generating misleading content ❌ Any exposure to minors — this is an 18+ model

Model Variants & Downloads

Hardware Requirements

Ethical Considerations

Bias & Fairness

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

Known Biases:

• —Single-character design reflects the creator's authored persona
• —Training data is small and limited in diversity (one persona, English)
• —Character design may perpetuate gender stereotypes

Mitigation:

• —Consider context before deployment
• —Validate outputs for bias
• —Supplement with diverse data if generalizing
• —Document known limitations

Safety & Responsible Use

Safety Features:

• —The character is designed around playful, consensual adult roleplay — not aggression
• —Reasoning channel makes the model's intent inspectable before it answers

Recommendations:

• —Gate access (this repo is gated) and verify users are adults
• —Use content filtering for any public deployment
• —Keep human oversight for anything beyond entertainment
• —Document limitations to users

Privacy & Data

• —Training data: private, proprietary dataset
• —No personal data in the training set
• —No data collection at inference

Terms of Service

By using this model, you agree to:

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

Third-Party Components

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

Maintenance & Support

Model Status

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

Support & Contact

Primary Contact: devops.ross@gmail.com

Organization:

• —Name: Ross Technologies AI Research Team
• —Email: devops.ross@gmail.com

Developer Resources:

• —Personal Site: https://ross-developers.com
• —GitHub: https://github.com/ross-sec
• —Model Hub: https://huggingface.co/ross-dev

Reporting Issues

1. 1.Email: devops.ross@gmail.com (include full details)
2. 2.Hugging Face: leave a comment on the model card

Response Time: best effort basis

Citation & Attribution

@model{sexygpt_v3_thinking_female_2026,
  title={SexyGPT-v3-Thinking-Female: Teaching a 27B Model to Wear a Character via SFT, Reward Modeling, and GRPO},
  author={Ross, Andrei},
  organization={Ross Technologies AI Research Team},
  year={2026},
  howpublished={\url{https://huggingface.co/ross-dev/SexyGPT-v3-Thinking-Female}}
}

Acknowledgments

• —Alibaba Qwen Team: for the Qwen3.8 base model and its thinking capabilities
• —Unsloth: for the QLoRA / GRPO training stack on consumer GPUs
• —Hugging Face: for the hub and transformers/TRL libraries
• —Author: Andrei Ross

Contact Information

📧 Email: devops.ross@gmail.com

🌐 Website:

• —https://ross-developers.com

💻 GitHub: https://github.com/ross-sec

Developer: Andrei Ross — Lead Developer (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