ORP-0.8B GGUF

This repository provides GGUF quants for hamzah0asadullah/ORP-0.8B.

ORP (Old Role Play) 0.8B is a full-parameter SFT fine-tune of Qwen/Qwen3.5-0.8B on the PygmalionAI/PIPPA conversation archive. It is trained to produce concise, responsive dialogue in the style of classic conversational roleplay systems while maintaining coherence on compact hardware.

All GGUF files were converted from the original safetensors weights using llama.cpp at native BF16 precision, then quantized into standard k-quant variants.

---

Available Files and Quantizations

---

Requirements

ORP-0.8B uses the Qwen 3.5 architecture (Qwen3_5ForConditionalGeneration / qwen35 in llama.cpp), which incorporates hybrid linear attention and interleaved MRoPE.

Make sure your local llama.cpp build or downstream interface (LM Studio, Ollama, Jan) includes Qwen 3.5 support.

---

How to Use

1. With llama.cpp

Run the model from the terminal using llama-cli:

llama-cli -m ./ORP-0.8B-Q4_K_M.gguf \
    -p "<|im_start|>system\nYou are a helpful roleplay assistant.<|im_end|>\n<|im_start|>user\nHey, how have you been?<|im_end|>\n<|im_start|>assistant\n" \
    -n 256 \
    -c 4096 \
    --temp 0.7

To run a local OpenAI-compatible API server:

llama-server -m ./ORP-0.8B-Q4_K_M.gguf \
    --host 127.0.0.1 \
    --port 8080 \
    -c 4096

2. With Ollama

Create a file named Modelfile in the same directory:

FROM ./ORP-0.8B-Q4_K_M.gguf

TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7

Then register and launch the model:

ollama create orp-0.8b -f Modelfile
ollama run orp-0.8b

3. With LM Studio

1. Copy or move any .gguf file from this folder into your LM Studio models directory.
2. Search for ORP-0.8B in your local models tab.
3. Load the model and select ChatML as the prompt format preset.

---

Model Details

• Original weights: hamzah0asadullah/ORP-0.8B
• Base model: Qwen/Qwen3.5-0.8B
• Training dataset: PygmalionAI/PIPPA
• Architecture: Qwen 3.5 (Qwen3_5ForConditionalGeneration)
• Parameters: ~0.85B
• Context window: Up to 262,144 tokens
• Vocabulary size: 248,320 tokens
• License: Apache 2.0

---

Attribution & Credits

Original model fine-tuning by hamzah0asadullah. Base model by Qwen. Training dataset by PygmalionAI.

GGUF conversions produced by Pink Pixel.

Made with 💖 by Pink Pixel