base_model:

• Qwen/Qwen3.8-27B
• orcarouter/Qwen3.8-27B-Uncensored
  library_name: transformers
  tags:
• qwen3
• lora
• merged
• roleplay
• persona
• conversational
• bilingual
  language:
• id
• en
  license: apache-2.0

Qwen3.8-27B-Uncensored-Impulsive

This model is a merge of a LoRA adapter into the base model orcarouter/Qwen3.8-27B-Uncensored in full BF16 precision. The fine-tuning focuses on a strong character persona and a bilingual conversational style, in Indonesian and English.

Status: experimental. This model has not been formally evaluated. Early tests show the persona and context recall working, but consistency across long conversations and across both languages has not been thoroughly tested.

Training Details

• Base model: orcarouter/Qwen3.8-27B-Uncensored
• Method: LoRA, r=32, lora_alpha=64, lora_dropout=0.05, use_rslora=False
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Precision: native BF16, no quantization
• Hyperparameters: 2 epochs, learning rate 1e-4, cosine scheduler, 5% warmup, weight decay 0.01, effective batch size 2
• Data format: Qwen ChatML (<|im_start|> / <|im_end|>)
• Loss: completion-only, computed only on assistant response tokens. System and user tokens are masked.
• Sample length: up to 8192 tokens. Longer conversations are split at message boundaries, so no tokens are discarded.
• Dataset: about 15 long bilingual conversations, roughly 50% Indonesian and 50% English, totaling about 550K tokens, averaging about 37K tokens per conversation. Each conversation has a strong persona and rich context.

Usage

The model uses the Qwen3 chat format. Use a system prompt to set the persona.

Example with vLLM as an OpenAI-compatible server:

vllm serve Sorenvza/Qwen3.8-27B-Uncensored-Impulsive \
  --dtype bfloat16 \
  --max-model-len 16384 \
  --reasoning-parser qwen3 \
  --api-key <YOUR_API_KEY>

Example request:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="<YOUR_API_KEY>")
resp = client.chat.completions.create(
    model="Sorenvza/Qwen3.8-27B-Uncensored-Impulsive",
    messages=[
        {"role": "system", "content": "You are a beautiful, fierce blonde character."},
        {"role": "user", "content": "Hi! Introduce yourself."},
    ],
    temperature=0.7,
    top_p=0.9,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(resp.choices[0].message.content)

Limitations

• Limited variety. Although the total token count is large, the data comes from about 15 characters and scenarios. The model may stay close to the training examples and may not generalize the persona to other situations.
• Persona consistency is not locked. Character name and traits may change between sessions unless the system prompt fixes those details.
• Reasoning leakage. The model tends to produce <think> text at the start of answers. Use the reasoning parser on the server, or disable thinking with enable_thinking: False.
• Long-context inference. Training conversations can be very long, but performance on long contexts during inference has not been validated.
• Chunks without a system prompt. Some training samples are pieces of long conversations that do not start with a system prompt, so the persona in those pieces may be weaker.
• Language balance. The data is roughly balanced between Indonesian and English. Performance in each language has not been evaluated separately.

Content and Use

This model has no additional content filtering on top of the base model. Outputs may include content that is inappropriate, sensitive, or inaccurate. Users are responsible for how they use and distribute it, including compliance with the laws that apply in their jurisdiction.

Do not use this model to target, harass, or deceive others, or to produce content involving minors in any form.