qwen3-companion-lora

Qwen3 Companion LoRA Adapter

Fine-tuned LoRA adapter for Qwen3-4B-Instruct model, trained for empathetic conversational interactions.

Model Details

• Base Model: Qwen/Qwen3-4B-Instruct-2507
• LoRA Rank: 32
• LoRA Alpha: 64
• Training Data: ~6,200 conversations (~20,000 examples)
• Training Metrics: 94%+ token accuracy, 0.19-0.22 loss

Usage

With vLLM (RunPod)

# Use with vLLM's LoRA support
--enable-lora \
--lora-modules companion=leiox777/qwen3-companion-lora

With Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
model = PeftModel.from_pretrained(model, "leiox777/qwen3-companion-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")

# Use for inference
messages = [
    {"role": "system", "content": "You are a supportive and empathetic conversational companion."},
    {"role": "user", "content": "Hello!"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))

Training Configuration

• Optimizer: AdamW (8-bit)
• Learning Rate: 2e-4
• Batch Size: 2 (gradient accumulation: 4)
• Epochs: 3
• Loss Function: Cross-entropy (assistant-only)

License

Apache 2.0