Skip to content

Share Your Vectors and Datasets

Contributing to the PSYCTL community is straightforward. Extract a vector or generate a dataset, upload to HuggingFace Hub, and submit a pull request to register it.

Share a Steering Vector

Step 1: Extract your vector

Use any of the three extraction methods:

psyctl extract.steering \
  --model "meta-llama/Llama-3.1-8B-Instruct" \
  --layer "model.layers[13].mlp.down_proj" \
  --dataset "./my_dataset/data.jsonl" \
  --output "./my_vector.safetensors" \
  --method bipo

Or use the 04_extract_vector notebook for an interactive walkthrough.

Step 2: Upload to HuggingFace Hub

from huggingface_hub import HfApi

api = HfApi()

# Create a new model repository
api.create_repo("your-username/steering-vector-agreeableness", repo_type="model")

# Upload your vector file
api.upload_file(
    path_or_fileobj="./my_vector.safetensors",
    path_in_repo="steering_vector.safetensors",
    repo_id="your-username/steering-vector-agreeableness",
    repo_type="model",
)

Step 3: Register in the community

Open a pull request on GitHub adding your vector to docs/COMMUNITY.DATASETS.md.

Naming convention: {username}/steering-vector-{personality}-{language}

Share a Steering Dataset

Step 1: Generate your dataset

# With OpenRouter API (no GPU needed)
psyctl dataset.build.steer \
  --model "qwen/qwen3-next-80b-a3b-instruct" \
  --personality "Agreeableness" \
  --output "./my_dataset" \
  --limit-samples 100

# Or with a local model
psyctl dataset.build.steer \
  --model "google/gemma-3-270m-it" \
  --personality "Agreeableness" \
  --output "./my_dataset" \
  --limit-samples 100

Or use the 03_generate_dataset notebook.

Step 2: Upload to HuggingFace Hub

from psyctl.core.dataset_builder import DatasetBuilder

builder = DatasetBuilder()
repo_url = builder.upload_to_hub(
    jsonl_file="./my_dataset/caa_dataset_*.jsonl",
    repo_id="your-username/steer-personality-agreeableness-en",
    private=False,
    license="mit",
)
print(f"Uploaded: {repo_url}")

Step 3: Register in the community

Open a pull request adding your dataset to docs/COMMUNITY.DATASETS.md.

Naming convention: {username}/steer-personality-{trait}-{language}

Dataset format (JSONL):

{
  "situation": "Context and dialogue...",
  "char_name": "Character name",
  "positive": "Response exhibiting the target personality",
  "neutral": "Neutral baseline response"
}

Tips for high-quality contributions

  • More samples = better vectors. 100+ samples recommended, 500+ for best results.
  • Use diverse source data. The allenai/soda dataset covers many social situations.
  • Test before sharing. Run a quick benchmark with psyctl benchmark inventory to verify your vector actually shifts the target personality.
  • Include metadata. Document the model, method, layer, and personality in your HuggingFace repo README.
  • License clearly. MIT or CC-BY recommended for maximum reusability.