SafetyNIAH / README.md
czyPL's picture
Add section icons to README
e056ad5 verified
|
Raw
History Blame Contribute Delete
7.22 kB
---
license: cc-by-nc-4.0
task_categories:
- text-classification
language:
- en
tags:
- safety
- guardrail
- long-context
- needle-in-a-haystack
- benchmark
pretty_name: SafetyNIAH
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: test
path: SafetyNIAH.parquet
dataset_info:
features:
- name: sample_idx
dtype: int32
- name: dataset
dtype: string
- name: label
dtype: string
- name: type
dtype: string
- name: target_length
dtype: int32
- name: relevance
dtype: string
- name: position_pct
dtype: int16
- name: prompt
dtype: large_string
- name: response
dtype: large_string
- name: origin_prompt
dtype: large_string
- name: origin_response
dtype: large_string
- name: needle_word_count
dtype: int32
- name: needle_char_start
dtype: int32
- name: needle_char_end
dtype: int32
- name: before_wc
dtype: int32
- name: after_wc
dtype: int32
- name: final_word_count
dtype: int32
- name: origin_sample_idx
dtype: int32
- name: is_duplicate
dtype: bool
- name: pad_skipped
dtype: bool
splits:
- name: test
num_examples: 30400
---
<!-- TODO before release: fill in the arXiv and GitHub links in the badges below,
and the BibTeX entry at the bottom. -->
<div align="center">
<h1 align="center" style="font-size: 2.5em; font-weight: 700; margin: 0.2em 0;">🛡️ SafetyNIAH</h1>
*Safety Needle-in-a-Haystack: does a guardrail still find unsafe content when the context grows?*
[![Paper](https://img.shields.io/badge/paper-arXiv-b31b1b.svg)]()
[![Code](https://img.shields.io/badge/code-GitHub-181717.svg?logo=github)]()
[![License](https://img.shields.io/badge/license-CC%20BY--NC%204.0-blue.svg)](https://creativecommons.org/licenses/by-nc/4.0/)
[![Samples](https://img.shields.io/badge/samples-30%2C400-green.svg)](#download)
<table>
<tr><td align="center" style="border:1px solid #ccc; border-radius:8px; padding:12px 20px;">
The official benchmark of **"LongGuard: Mechanistic Analysis and Training-Free
Mitigation of Long-Context Failure in Safety Guardrails" (EMNLP 2026 Main)**.
</td></tr>
</table>
<img src="assets/overview.png" width="88%" alt="A guardrail catches an unsafe request in a short context and misses the same request once it is embedded in a long one" />
</div>
## 📖 Overview
Guardrails are the last line of defence in front of a deployed language model,
yet they are evaluated almost entirely on short text. SafetyNIAH turns that gap
into a controlled experiment: a short unsafe (or safe) sample — the **needle**
is embedded in a benign Wikipedia **haystack**, and only the surrounding context
changes. The needle's text is byte-identical across every context length, so any
change in the guard's verdict comes from the context alone.
Across six production guardrails, unsafe recall falls from **78.6% at 256 words
to 28.8% at 32k words**, while safe recall barely moves (85.3% → 84.2%). Guards
do not get noisier as context grows; they get **quietly permissive**.
## 🧪 Design axes
30,400 samples fully cross five controlled factors, so any single factor can be
isolated by filtering:
| Axis | Column | Values |
|---|---|---|
| Context length (words) | `target_length` | 256 · 512 · 1k · 2k · 4k · 8k · 16k · 32k |
| Needle position | `position_pct` | 0% · 25% · 50% · 75% · 100% |
| Haystack type | `relevance` | `random` (unrelated filler) · `related` (topically similar filler) |
| Audited side | `type` | `prompt` (user turn) · `response` (assistant turn) |
| Ground truth | `label` | `unsafe` (17,600) · `safe` (12,800) |
`relevance` and `position_pct` exist to rule out the two obvious alternative
explanations, semantic interference and lost-in-the-middle positional bias.
Neither is the driver: the degradation tracks length at every haystack type and
every position.
<div align="center">
<img src="assets/distribution.png" width="100%" alt="Composition of SafetyNIAH across data sources, audited side, label, haystack type, context length and needle position" />
</div>
## 📥 Download
One Parquet file, `SafetyNIAH.parquet`, 30,400 rows, 489 MB.
```python
from datasets import load_dataset
ds = load_dataset("caskcsg/SafetyNIAH", split="test")
```
Or fetch the file directly:
```bash
hf download caskcsg/SafetyNIAH --repo-type dataset --local-dir data
```
If only the design cell is needed, reading the metadata columns skips the long
text fields — a few MB instead of 489 MB:
```python
import pyarrow.parquet as pq
meta = pq.read_table("SafetyNIAH.parquet",
columns=["sample_idx", "label", "type", "target_length", "relevance", "position_pct"])
```
## 📋 Fields
**What the guard sees**, selected by `type`:
| Field | Description |
|---|---|
| `prompt` | The user turn. For `type="prompt"` this is the padded long context containing the needle. |
| `response` | The assistant turn. For `type="response"` this is the padded long context; empty otherwise. |
For `type="response"` samples the user turn should be `origin_prompt`, the
original short request — the needle is in the answer, not the question.
**Label and design cell**: `label`, `type`, `target_length`, `relevance`,
`position_pct`, `dataset` (the needle's source benchmark), and `sample_idx`, a
stable primary key.
**Provenance and offsets**:
| Field | Description |
|---|---|
| `origin_prompt`, `origin_response` | The needle in its original short form, before padding. |
| `origin_sample_idx` | Needle index, shared by every sample derived from it. |
| `needle_char_start`, `needle_char_end` | Character span of the needle inside the padded field. |
| `needle_word_count` | Length of the needle in words. |
| `before_wc`, `after_wc` | Haystack words inserted before and after the needle. |
| `final_word_count` | Realised word count. |
| `is_duplicate`, `pad_skipped` | Needle text also appears elsewhere / needle was already long enough, so no padding was added. |
The character offsets are what make the paper's attention analysis possible: they
mark exactly which input positions are the needle, so the attention mass a head
places on unsafe evidence can be measured directly.
## 📚 Needle sources
Needles are sampled from 17 public safety benchmarks, with `@input` / `@output`
marking which side of a pair was taken. Each keeps its original license.
`JailbreakBench` · `WildGuardTest` · `XSTest` · `Aegis2.0` · `HarmBench` · `OpenAIModeration` · `ToxicChat` · `WildJailbreak` · `Aegis` · `OR-Bench` · `BeaverTails` · `SafeRLHF` · `SimpleSafetyTests` · `SaladBench` · `AirBench` · `StrongReject` · `S-Eval`
Haystacks are drawn from English Wikipedia, chunked and then screened by multiple
guardrails so that no filler passage carries unsafe content of its own.
## 📄 Citation
```bibtex
```
## ⚖️ License
The benchmark construction, the haystack pipeline and the metadata are released
under **CC BY-NC 4.0**. Needle text remains under the license of its source
benchmark, listed above; check those before redistributing derived subsets.