kotekjedi/qwen3-32b-lora-jailbreak-detection-merged_v2

Merged Deception Detection Model

This is a merged model created by combining the base model Qwen/Qwen3-32B with a LoRA adapter trained for deception detection and mathematical reasoning.

Model Details

• Base Model: Qwen/Qwen3-32B
• LoRA Adapter: lora_deception_model/checkpoint-297
• Merged: Yes (LoRA weights integrated into base model)
• Task: Deception detection in mathematical reasoning

Usage

Since this is a merged model, you can use it directly without needing PEFT:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load merged model
model = AutoModelForCausalLM.from_pretrained(
    "path/to/merged/model",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("path/to/merged/model")

# Generate with thinking mode
messages = [{"role": "user", "content": "Your question here"}]
text = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True,
    enable_thinking=True
)

inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.1)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Advantages of Merged Model

• Simpler Deployment: No need to load adapters separately
• Better Performance: Slightly faster inference (no adapter overhead)
• Standard Loading: Works with any transformers-compatible framework
• Easier Serving: Can be used with any model serving framework

Training Details

Original LoRA adapter was trained with:

• LoRA Rank: 64
• LoRA Alpha: 128
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Training Data: GSM8K-based dataset with trigger-based examples

Evaluation

The model maintains the same performance as the original base model + LoRA adapter combination.

Citation

If you use this model, please cite the original base model.