This repository contains a PEFT LoRA adapter for Qwen/Qwen3.8-27B, released as Erato in the Presence Muses series.
It is a Direct Preference Optimization (DPO) adapter for roleplay / companion chat with send_message tool calls.
The base model is not included. Load the adapter on top of
Qwen/Qwen3.8-27B.
| Item | Value |
|---|---|
| Developed by | PresenceAI |
| Series | Presence Muses |
| Release | v1.0 (GA) |
| Model type | LoRA adapter (PEFT), DPO |
| Training stage | Direct Preference Optimization (DPO) |
| Base model | Qwen/Qwen3.8-27B |
| Languages | English, Chinese |
| License | Apache-2.0 |
| Adapter rank | r = 32, alpha = 32, dropout = 0.0 |
| Target modules | Attention and MLP projections matched by the PEFT regex in adapter_config.json |
Qwen/Qwen3.8-27B ──DPO──▶ Erato-27B-DPO-LoRA-v1.0 (this repository)
Initialized directly from the base model and aligned with DPO. No intermediate checkpoint or upstream adapter is required — load this adapter on Qwen/Qwen3.8-27B as shown below. The base weights are not included in this repository.
send_message call.Out of scope: factual question answering, medical / legal / financial advice, high-stakes decision making.
chosen response is a human-written reply; rejected is a model-generated alternative. The preference data is not released.adapter_config.json for target modules).The adapter package does not include a complete DPO training-argument dump. Refer to the upstream base model requirements when reproducing training.
No automatic benchmark or human preference result is included in this release. Evaluate on your own held-out conversations before deployment.
The base model uses the Qwen3.5 architecture (Qwen3_5ForConditionalGeneration). Use a recent Transformers release that supports it (transformers >= 5.16 in our deployment) and a compatible PEFT version.
For vLLM tool-calling, the deployment used the qwen3_xml tool-call parser.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_id = "Qwen/Qwen3.8-27B"
adapter_id = "PresenceAI/Erato-27B-DPO-LoRA-v1.0"
tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(
base_id, torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()
vllm serve Qwen/Qwen3.8-27B \
--enable-lora --max-lora-rank 32 \
--lora-modules erato=PresenceAI/Erato-27B-DPO-LoRA-v1.0 \
--enable-auto-tool-choice --tool-call-parser qwen3_xml
Use the tokenizer and chat_template.jinja included in this repository to keep the prompt and tool-call format consistent with training.
The adapter is released under the Apache License 2.0. The base model Qwen/Qwen3.8-27B is Apache-2.0. See LICENSE and NOTICE.
@misc{presenceai_erato_27b_dpo_lora_v1,
title = {PresenceAI Erato-27B-DPO-LoRA v1.0: roleplay and tool-calling chat},
author = {PresenceAI},
year = {2026},
note = {LoRA adapter, Apache-2.0}
}
Qwen/Qwen3.8-27B