arda24/humanizer-final

Humanizer LoRA Adapter

This is a LoRA (Low-Rank Adaptation) adapter for Llama3 8B Instruct that converts formal text into more natural, human-like language.

Model Details

• Base Model: meta-llama/Meta-Llama-3-8B-Instruct
• Adapter Type: LoRA (Low-Rank Adaptation)
• LoRA Rank: 32
• LoRA Alpha: 64
• Target Modules: {'k_proj', 'gate_proj', 'q_proj', 'v_proj', 'up_proj', 'down_proj', 'o_proj'}
• Task: Text humanization - converting formal/academic text to conversational style

Files Included

This adapter includes all necessary files:

• adapter_config.json - LoRA configuration
• adapter_model.safetensors - LoRA weights
• special_tokens_map.json - Special tokens mapping
• tokenizer.json - Tokenizer vocabulary
• tokenizer_config.json - Tokenizer configuration
• training_args.bin - Training arguments

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and tokenizer
base_model = "meta-llama/Meta-Llama-3-8B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Load LoRA adapter
adapter_name = "arda24/Humanizer"
model = PeftModel.from_pretrained(model, adapter_name)

# Prepare input
prompt = "### Instruction:
rewrite this text in a natural and human like way

### Input:
The system requires authentication before proceeding.

### Response:
"

# Generate humanized text
inputs = tokenizer(prompt, return_tensors="pt")
if torch.cuda.is_available():
    inputs = {k: v.cuda() for k, v in inputs.items()}

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.3,
    do_sample=True,
    top_p=0.7,
    repetition_penalty=1.05,
    no_repeat_ngram_size=2
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
humanized_text = response.split("### Response:")[1].strip()
print(humanized_text)

Example

Input: "The system requires authentication before proceeding."

Output: "You need to log in first before you can access the system."

Training Configuration

• LoRA Rank: 32
• LoRA Alpha: 64
• Learning Rate: 1e-5
• Batch Size: 1
• Gradient Accumulation Steps: 16
• Training Steps: ~4000

Advantages of LoRA

• Smaller size: Only ~50MB vs several GB for full model
• Faster loading: Loads quickly on top of base model
• Flexible: Can be combined with other adapters
• Efficient: Uses minimal additional parameters

Limitations

• Works best with formal/academic text
• May occasionally add citations if not properly controlled
• Conservative settings recommended for minimal changes
• Not suitable for creative writing or fiction

License

This adapter is based on Llama3 8B Instruct and follows the same license terms.