license: apache-2.0
task_categories:
- text-generation
language: - en
tags: - activation-steering
- concept-steering
- flow-matching
- axbench
pretty_name: FLAS-Concept-46k
size_categories: - 1M<n<10M
FLAS-Concept-46k
A model-agnostic concept-steering corpus used to train FLAS (Flow-based Activation Steering). 46,472 unique concepts, ~2.64 M rows (concept × prompt × steered response).
- 🌐 Project page: https://flas-ai.github.io
- 📄 Paper: https://arxiv.org/abs/2605.05892
- 💻 Code: https://github.com/flas-ai/FLAS
What's in it
Each row is a (prompt, steered response, concept) triple:
| Column | Description |
|---|---|
input |
instruction / prompt |
output |
response exhibiting the concept |
concept |
natural-language concept text |
concept_id |
stable content hash (blake2b) — identical concepts collapse across source models |
concept_genre |
concept category |
dataset_category |
provenance split |
source_model |
which base-model concept set the concept came from |
The data is text-only and model-agnostic — FLAS extracts activations live at its chosen layer, so this one corpus trains a steerer for any base model.
How it was built
Built on AxBench Concept16k_v2 (pyvene/axbench-concept16k_v2), integrating the three base-model concept splits — Gemma-2-2B, Gemma-2-9B, and Llama-3.1-8B — into one set of 46,472 unique concepts (deduplicated by content hash). The original gemma-2 splits were first released in pyvene/axbench-concept16k; v2 adds the Llama-3.1 split.
Because the AxBench concepts came with too few distinct prompts relative to the number of concepts (a handful of prompts were reused thousands of times — up to ~7,000× each), we augmented with ~30 additional prompts per concept, drawn from a filtered subset of Dolly-15k, and capped per-prompt reuse at 500 to break heavy-prompt dominance. Final corpus: ~2.64 M rows after dropping nulls/empties/refusals.
A stratified 100-sample C/I/F quality check (Azure GPT-4o-mini judge) scored the corpus at HMean ≈ 1.38 (AxBench-side 1.35, augmented-side 1.41), confirming the augmented prompts match the original split's quality.
Files
train.parquet— the corpus (~887 MB).train.metadata.json— build provenance (source files, dedup/cap stats, quality check).
Citation
@article{flas2026,
title={Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Intervention},
author={Zehao Jin and Ruixuan Deng and Junran Wang and Xinjie Shen and Chao Zhang},
year={2026},
eprint={2605.05892},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2605.05892},
}
Built on AxBench (Wu et al., 2025) — github.com/stanfordnlp/axbench — and Databricks Dolly-15k.