model_usage.md

Using SPA from Hugging Face Hub

This guide shows you how to use Selective Prompt Anchoring (SPA) directly from the Hugging Face Hub.

Installation

First, install the required packages:

pip install huggingface_hub torch transformers

Basic Usage

Here's a simple example of how to use SPA:

from huggingface_hub import hf_hub_download
import sys
import os

# Download the SPA files from the Hub
repo_id = "YOUR_USERNAME/selective-prompt-anchoring"  # Replace with your username
spa_file = hf_hub_download(repo_id=repo_id, filename="spa.py")
spa_hf_file = hf_hub_download(repo_id=repo_id, filename="spa_hf.py")

# Add the directory containing the files to the Python path
sys.path.append(os.path.dirname(spa_file))

# Now you can import from the downloaded files
from spa_hf import load_spa_model

# Load SPA with just the base model, ignoring any saved configurations
spa_model = load_spa_model(
    model_name=None,  # Skip loading from Hub
    base_model_name="Qwen/Qwen3-0.6B",  # Use your preferred base model
    device_map="auto"  # Use "cpu" if you don't have a GPU
)

# Use SPA to emphasize specific tokens
prompt = "How is the weather today?"
anchors = ["today"]

output = spa_model.generate_with_spa(
    prompt=prompt,
    anchors=anchors,
    anchoring_strength=3.0,  # Adjust strength to control anchor influence
    max_new_tokens=200
)

print(output)

Advanced Usage

For more advanced usage, including different input formats and parameter settings, check out the GitHub repository or the Documentation.