Datasetpublic

seoirsem/tulu3-SFT-500k-25k-data-attributes-full

SURF Attributes (Full) Complete dataset for SURF research and extension. Quick Start For running SURF, use the minimal dataset: seoirsem/tulu3-SFT-500k-25k-data-attributes uv run -m surf.cli.main sweep \ --attributes seoirsem/tulu3-SFT-500k-25k-data-attributes \ --rubric rubrics/rebuttal.yaml \ -o results/ Dataset Fields prompt: The query text response: The model response (if available) attributes: Raw extracted attributes (10 per… See the full description on the dataset page: https://huggingface.co/datasets/seoirsem/tulu3-SFT-500k-25k-data-attributes-full.

sourceHugging Faceupdated 8mo agoView on Hugging Face

0likes39downloads

Dataset Card

SURF Attributes (Full)

Complete dataset for SURF research and extension.

Quick Start

For running SURF, use the minimal dataset: seoirsem/tulu3-SFT-500k-25k-data-attributes

bash

uv run -m surf.cli.main sweep \
    --attributes seoirsem/tulu3-SFT-500k-25k-data-attributes \
    --rubric rubrics/rebuttal.yaml \
    -o results/

Dataset Fields

—prompt: The query text
—response: The model response (if available)
—attributes: Raw extracted attributes (10 per query)
—sae_attributes: Semantic cluster summaries
—source: Source dataset identifier

Additional Files

centroids.npy

Cluster centroids for embedding-based attribute lookup.

—Shape: (25000, 4096)
—Embedding model: Qwen/Qwen3-Embedding-8B

python

import numpy as np
from huggingface_hub import hf_hub_download

centroids_path = hf_hub_download(
    repo_id="seoirsem/tulu3-SFT-500k-25k-data-attributes-full",
    filename="centroids.npy",
    repo_type="dataset"
)
centroids = np.load(centroids_path)

cluster_summaries.jsonl

Human-readable summaries for each cluster.