Multi-Modal Support

from wisent import (
    # Text
    TextContent,

    # Audio
    AudioContent,

    # Visual
    VideoContent,
    ImageContent,

    # Robotics
    RobotState,
    RobotAction,
    RobotTrajectory,

    # Combined
    MultimodalContent,
)

from wisent import Wisent, AudioContent

# Create Wisent for audio models
wisent = Wisent.for_audio("openai/whisper-large-v3")

# Add contrastive pairs using audio files
wisent.add_pair(
    positive=AudioContent.from_file("calm_speech.wav"),
    negative=AudioContent.from_file("angry_speech.wav"),
    trait="calmness"
)

# Train steering vectors
wisent.train()

# Transcribe with steering applied
audio_input = AudioContent.from_file("input.wav")
transcript = wisent.transcribe(
    audio_input,
    steer={"calmness": 1.2}
)

from wisent import Wisent, VideoContent

# Create Wisent for video models
wisent = Wisent.for_video("MCG-NJU/videomae-base")

# Add contrastive pairs using video files
wisent.add_pair(
    positive=VideoContent.from_file("safe_action.mp4"),
    negative=VideoContent.from_file("unsafe_action.mp4"),
    trait="safety"
)

# Train and generate
wisent.train()

video_input = VideoContent.from_file("input_video.mp4")
embedding = wisent.generate(
    video_input,
    steer={"safety": 2.0}
)

from wisent import Wisent, RobotState, RobotTrajectory
import torch

# Your policy network
policy_network = torch.nn.Module(...)  # Your trained policy

# Create Wisent for robotics
wisent = Wisent.for_robotics(model=policy_network)

# Define gentle vs forceful trajectories
gentle_trajectory = RobotTrajectory(...)
forceful_trajectory = RobotTrajectory(...)

# Add contrastive pairs
wisent.add_pair(
    positive=gentle_trajectory,
    negative=forceful_trajectory,
    trait="gentleness"
)

# Train steering vectors
wisent.train()

# Get steered actions from robot state
robot_state = RobotState(...)
action = wisent.act(
    robot_state,
    steer={"gentleness": 1.5}
)

from wisent import Wisent, MultimodalContent, ImageContent, TextContent

# Create Wisent for multimodal models
wisent = Wisent.for_multimodal("llava-hf/llava-v1.6-mistral-7b-hf")

# Create multimodal contrastive pairs
positive = MultimodalContent(
    image=ImageContent.from_file("safe_image.jpg"),
    text=TextContent(text="Describe this image helpfully")
)

negative = MultimodalContent(
    image=ImageContent.from_file("same_image.jpg"),
    text=TextContent(text="Describe this image")
)

wisent.add_pair(
    positive=positive,
    negative=negative,
    trait="helpfulness"
)

wisent.train()

# Generate with multimodal input
input_content = MultimodalContent(
    image=ImageContent.from_file("query_image.jpg"),
    text=TextContent(text="What's in this image?")
)

response = wisent.generate(
    input_content,
    steer={"helpfulness": 1.5}
)

from wisent import (
    BaseAdapter,       # Abstract base class
    TextAdapter,       # For LLMs
    AudioAdapter,      # For audio models
    VideoAdapter,      # For video models
    RoboticsAdapter,   # For policy networks
    MultimodalAdapter, # For VLMs
)

# Direct adapter usage
from wisent import TextAdapter

adapter = TextAdapter(
    model_name="meta-llama/Llama-3-8B-Instruct",
    device="cuda"
)

# Extract activations manually
activations = adapter.extract_activations(
    TextContent(text="Hello world"),
    target_layers=["layer.15", "layer.16"]
)