from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model - ITI automatically applies!
model = AutoModelForCausalLM.from_pretrained(
    "YOUR_USERNAME/llama-31-8b-creativity-iti",
    trust_remote_code=True,  # Required for auto-ITI
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/llama-31-8b-creativity-iti")

# Generate creative code
prompt = "Write a function to check if a number is prime"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.8)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Configuration

Alpha: 0.15
Active Heads: 48
Base Model: LLaMA 3.1 8B Instruct
Intervention: Automatic during inference

How It Works

The model automatically applies Inference-Time Intervention to enhance creativity:

Monitors 48 attention heads during generation
Shifts activations by α=0.15 toward creative directions
Results in more innovative code solutions

Training

Dataset: NeoCoder (1058 problems)
Method: Extracted activations from complete solutions
Metric: Novel technique usage vs human solutions

License

Apache 2.0

syed-aliredha/llama-31-8b-creativity-it-60-percent

LLaMA 3.1 8B with Creativity ITI

Quick Start

Configuration

How It Works

Training

License

syed-aliredha/
llama-31-8b-creativity-it-60-percent