How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf RX5950XT/silicon-based-girlfriend:Q8_0
# Run inference directly in the terminal:
llama cli -hf RX5950XT/silicon-based-girlfriend:Q8_0
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf RX5950XT/silicon-based-girlfriend:Q8_0
# Run inference directly in the terminal:
llama cli -hf RX5950XT/silicon-based-girlfriend:Q8_0
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf RX5950XT/silicon-based-girlfriend:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf RX5950XT/silicon-based-girlfriend:Q8_0
Use Docker
docker model run hf.co/RX5950XT/silicon-based-girlfriend:Q8_0
Quick Links

Silicon-Based-Girlfriend โ€” QLoRA Adapter

๐Ÿš€ ้‡ๅคงๅ‡็ดš๏ผš็ŸฝๅŸบๅฅณๅ‹ v2๏ผˆSilicon-Based-Girlfriend v2๏ผ‰ๅทฒๆญฃๅผ็™ผๅธƒ๏ผ

ๅ‡็ดš่‡ณ 9B ๅŸบๅบ•๏ผˆHuihui-Qwen3.5-9B-abliterated๏ผ‰๏ผŒๆญท็ถ“ SFT ่ˆ‡ๅ…ฉ่ผช GRPO ๅผทๅŒ–ๅญธ็ฟ’่จ“็ทด๏ผŒๅ…จ้ข่งฃๆฑบ่ทจ่ผช้‡่ค‡ๅ•้กŒ๏ผŒๅคงๅน…ๆๅ‡ๅ…ฅๆˆฒๆ„Ÿ่ˆ‡ๅคš่ผช้€ฃ่ฒซๆ€ง๏ผš


ๅŸบๆ–ผ Qwen3.5-4B ็š„ QLoRA ๅพฎ่ชฟ Adapter๏ผŒ่จ“็ทด็›ฎๆจ™็‚บๆฒ‰ๆตธๅผ็น้ซ”ไธญๆ–‡่ง’่‰ฒๆ‰ฎๆผ”ใ€‚ๆœฌๅ€‰ๅบซๅŒ…ๅซ LoRA Adapter ๆฌŠ้‡่ˆ‡ GGUF ๆ ผๅผๆจกๅž‹ใ€‚


Model Details / ๆจกๅž‹่ณ‡่จŠ

้ …็›ฎ ๅ…งๅฎน
Base Model Qwen/Qwen3.5-4B
Fine-tuning Method QLoRA (4-bit NF4)
LoRA Rank 32
LoRA Alpha 64
LoRA Dropout 0.05
LoRA Target All linear layers
Training Epochs 5
Context Length 8192 tokens
Learning Rate 1e-4
LR Scheduler Cosine
Optimizer paged_adamw_8bit
Training Samples 985
Train Loss 1.108
Eval Loss 1.434
Hardware NVIDIA RTX A6000 (48GB VRAM)
Training Time ~19 hours
Framework LLaMA-Factory
Chat Template qwen3_5_nothink (non-thinking mode)

Files / ๆช”ๆกˆ่ชชๆ˜Ž

ๆช”ๆกˆ ่ชชๆ˜Ž
adapter_config.json LoRA ่จญๅฎšๆช”
adapter_model.safetensors LoRA ๆฌŠ้‡๏ผˆ248 MB๏ผ‰
tokenizer_config.json Tokenizer ่จญๅฎš๏ผˆๅซ nothink chat template๏ผ‰
tokenizer.json Tokenizer
vocab.json / merges.txt Vocabulary
silicon-gf-q8_0.gguf Q8_0 ้‡ๅŒ– GGUF๏ผˆ4.2 GB๏ผŒ้ฉ็”จ llama.cpp / LM Studio๏ผ‰
training_loss.png ่จ“็ทด Loss ๆ›ฒ็ทš
training_eval_loss.png ่ฉ•ไผฐ Loss ๆ›ฒ็ทš

Usage / ไฝฟ็”จๆ–นๅผ

Option 1: GGUF (Recommended / ๆŽจ่–ฆ)

็›ดๆŽฅๅœจ LM Studio ๆˆ– llama.cpp ่ผ‰ๅ…ฅ silicon-gf-q8_0.gguf๏ผŒ็„ก้œ€้กๅค–ๅฎ‰่ฃใ€‚

# llama.cpp
./llama-cli -m silicon-gf-q8_0.gguf -c 8192 --temp 0.8

Option 2: LoRA Adapter with transformers + PEFT

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3.5-4B"
adapter = "RX5950XTP/silicon-based-girlfriend"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

messages = [
    {"role": "user", "content": "ๅ˜ฟ๏ผŒไฝ ๅœจๅนนๅ˜›๏ผŸ"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.8, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Option 3: LLaMA-Factory inference

llamafactory-cli chat \
  --model_name_or_path Qwen/Qwen3.5-4B \
  --adapter_name_or_path RX5950XTP/silicon-based-girlfriend \
  --template qwen3_5_nothink \
  --finetuning_type lora

Training Curves / ่จ“็ทดๆ›ฒ็ทš

Training Loss Eval Loss


Dataset / ่จ“็ทด่ณ‡ๆ–™้›†

  • ๅ€‰ๅบซ๏ผšRX5950XTP/silicon-girlfriend-dataset
  • ๆ ผๅผ๏ผšShareGPT๏ผˆsystem + conversations with from/value๏ผ‰
  • ็ญ†ๆ•ธ๏ผš985 ็ญ†ๅคš่ผชๅฐ่ฉฑ
  • ่ชž่จ€๏ผš็น้ซ”ไธญๆ–‡๏ผˆ่‡บ็ฃ็”จ่ชž๏ผ‰
  • ็”Ÿๆˆๆ–นๅผ๏ผš็”ฑ Kimi K2.5 ๆ นๆ“š่ง’่‰ฒ่จญๅฎš็”Ÿๆˆ

Notes / ๆณจๆ„ไบ‹้ …

  • ๆœฌๆจกๅž‹ไฝฟ็”จ qwen3_5_nothink chat template๏ผŒ้ ่จญไธๅ•Ÿ็”จๆ€่€ƒๆจกๅผ๏ผŒๅ›ž่ฆ†ๆœƒ็›ดๆŽฅ่ผธๅ‡บ่ง’่‰ฒๅฐ่ฉฑใ€‚
  • ่ง’่‰ฒ่จญๅฎšๅŒ…ๅซไธ่‰ฏ็”จ่ชž่ˆ‡ๆˆไบบไธป้กŒ๏ผŒ่ซ‹่‡ช่กŒ่ฉ•ไผฐไฝฟ็”จๅ ดๆ™ฏใ€‚
  • ๆจกๅž‹ไปฅ QLoRA ่จ“็ทด๏ผŒๆŽจ็†ๆ™‚้œ€ๆญ้… base model๏ผˆQwen3.5-4B๏ผ‰ไธ€ๅŒ่ผ‰ๅ…ฅ๏ผŒๆˆ–็›ดๆŽฅไฝฟ็”จ GGUFใ€‚

License

Apache 2.0๏ผˆ้ตๅพช Qwen3.5-4B ๅŽŸๆŽˆๆฌŠ๏ผ‰

Downloads last month
193
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for RX5950XT/silicon-based-girlfriend

Finetuned
Qwen/Qwen3.5-4B
Adapter
(937)
this model

Dataset used to train RX5950XT/silicon-based-girlfriend