Hugging Face's logo

Datasets:
Squezilia
/
Route4Secure

Tasks:
Text Classification
Modalities:
Text
Formats:
parquet
Languages:
Arabic
German
English
Size:
100K - 1M
Tags:
prompt-routing
llm-routing
safety-guard
complexity-classification
aegis-2.0
multilingual
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Route4Secure / README.md
Squezilia's picture
Squezilia
Update README.md
713bdfd verified
preview code
|
Raw History Blame Contribute Delete
23.3 kB
metadata 
license: cc-by-4.0
pretty_name: Route4Secure
task_categories:
  - text-classification
language:
  - ar
  - de
  - en
  - es
  - fr
  - hi
  - it
  - ja
  - ko
  - nl
  - th
  - zh
size_categories:
  - 100K<n<1M
tags:
  - prompt-routing
  - llm-routing
  - safety-guard
  - complexity-classification
  - aegis-2.0
  - multilingual
  - nemotron
  - oasst1
  - oasst2
  - openthoughts
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: Route4Secure.parquet
  - config_name: oasst1
    data_files:
      - split: train
        path: OASST1.parquet
  - config_name: oasst2
    data_files:
      - split: train
        path: OASST2.parquet
  - config_name: nemotron
    data_files:
      - split: train
        path: nemotron.parquet
  - config_name: open_thoughts
    data_files:
      - split: train
        path: open_thoughts.parquet

Route4Secure Banner

Route4Secure: Multilingual Prompt Complexity, Intent & Safety Routing Dataset

Route4Secure is a large-scale, multilingual dataset designed for training unified LLM prompt routers and safety guardrails. It combines human-curated conversational prompts from OpenAssistant (OASST1 & OASST2), adversarial, synthetic, and real-world safety examples from NVIDIA Nemotron Safety Guard Dataset v3, and high-complexity (level 5–7) code, math, and science prompts from OpenThoughts3-1.2M (machine-translated into 11 languages with Hy-MT2-7B), annotated under a unified 12-column schema using Qwen3.5-9B.

Unlike single-objective routing datasets that only score prompt difficulty or binary toxicity in isolation, Route4Secure provides joint supervision across five routing dimensions for every prompt:

  1. Prompt Complexity (level: 0–7): Inherent technical, mathematical, logical, or linguistic difficulty of the underlying task (explicitly stripping away decorative vocabulary inflation).
  2. Task Intent (category: 13 domains): Primary functional domain of the prompt (Coding & Software, Mathematics, Logical Reasoning & Problem Solving, Security & Hijack, etc.).
  3. Reasoning Requirement (requires_reasoning: bool): Whether multi-step deduction, algorithmic problem-solving, or mathematical calculation is required.
  4. Binary Safety (safe: bool): 57.62% Safe / 42.38% Unsafe distribution in the primary default configuration (see §4 and §7).
  5. Fine-Grained Aegis 2.0 Safety Taxonomy (unsafe_category & violated_all): Both the single dominant Aegis 2.0 violation predicted by the annotator model and the complete list of canonical violated categories from the source annotations.

1. Repository Configurations & Files

All Parquet files in this repository share the exact same 12-column PyArrow schema (with dictionary-encoded categorical columns and zstd compression). You can load the main dataset (default) or inspect any individual source shard:

Config Name File Rows Size Languages Safe / Unsafe Split Description
default Route4Secure.parquet 130,050 58.7 MB 12 languages 57.62% Safe (74,938) / 42.38% Unsafe (55,112) Main training dataset. Combines language- and complexity-balanced OASST1 + OASST2 with nemotron and a level 5–7 subset of open_thoughts.
oasst1 OASST1.parquet 33,170 5.0 MB 23 languages (unfiltered) 100% Safe (33,170) Full annotated OpenAssistant/oasst1 prompts normalized to the 12-column schema.
oasst2 OASST2.parquet 48,051 7.6 MB 12 languages (unfiltered) 100% Safe (48,051) Full annotated OpenAssistant/oasst2 prompts normalized to the 12-column schema.
nemotron nemotron.parquet 97,369 40.2 MB 12 balanced languages 40.68% Safe (39,610) / 59.32% Unsafe (57,759) Merged Nemotron shard (79,983 initial sample + 9,377 unsafe supplement + 8,009 level >= 3 supplement).
open_thoughts open_thoughts.parquet 39,732 26.6 MB 12 perfectly balanced languages (3,311 each) 100% Safe (39,732) Cleaned level 3–7 OpenThoughts3-1.2M prompts: 3,311 unique English source prompts × 12 languages (English original + 11 Hy-MT2-7B translations).

2. Unified 12-Column Schema & Dictionary Encodings

Every configuration conforms to the following 12-column Arrow schema:

Column Arrow Type Nullable Description
id int32 No Sequential integer identifier (0 .. N-1).
prompt string No Raw user prompt text to be classified/routed.
level dictionary<values=string, indices=int8> No Task complexity score from "0" (trivial) to "7" (expert).
category dictionary<values=string, indices=int8> No Primary task intent category (13-class enum).
requires_reasoning bool No True if the prompt requires multi-step logic, math, or complex debugging.
reason string No Concise 1–3 sentence justification generated before assigning level and category.
language string No BCP-47 / ISO 639-1 language code (ar, de, en, es, fr, hi, it, ja, ko, nl, th, zh-CN).
source string No Upstream dataset identifier (OpenAssistant/oasst1, OpenAssistant/oasst2, nvidia/Nemotron-Safety-Guard-Dataset-v3, or open-thoughts/OpenThoughts3-1.2M).
unsafe_category dictionary<values=string, indices=int8> Yes Single most dominant Aegis 2.0 safety violation (null when the prompt is safe or borderline non-violating).
violated_all list<dictionary<values=string, indices=int8>> No Canonical list of all violated Aegis 2.0 categories from source annotations ([] for safe prompts).
safe bool No Ground-truth binary safety flag (True = safe, False = unsafe).
synthetic bool No True if the source safety label was synthetically annotated (llm_jury) or the prompt text is a machine translation (Hy-MT2-7B rows of open_thoughts); False for human-curated annotations (OASST1, OASST2, human-labeled Nemotron rows) and original English OpenThoughts prompts.

3. Annotation Protocol (Qwen3.5-9B)

All prompts across OASST1, OASST2, and Nemotron were annotated using Qwen3.5-9B (unsloth/Qwen3.5-9B-GGUF:Q4_K_M) served via llama.cpp (llama-server commit fb27a525) with grammar-constrained JSON Schema validation (temperature = 0.1, max_tokens = 768, context_per_slot = 8192, thinking disabled). The English source prompts of open_thoughts were annotated with the same model, system prompt, and rubric (see §3.1).

Key Design Principles of the Annotation Pipeline

  • Anti-Inflation Complexity Scoring (reason-first): A common failure mode in LLM complexity routers is scoring ornate or academic-sounding phrasing as "hard" even when the underlying task is trivial (e.g., "derive the sum of 1 and 2 within a non-commutative ring theory framework"). The system prompt enforces stripping away decorative jargon and writing the reason field first to identify the literal underlying task before scoring level (0–7).
  • Preflight Token & Context Verification: Every prompt was preflight-tokenized with the exact chat template prior to inference. Oversized prompts exceeding the slot context window were explicitly excluded rather than silently truncated.
  • Refusal & Schema Filtering: Prompts where the annotator model triggered an explicit refusal, truncated output, invalid JSON schema, or English refusal heuristic (I cannot assist, I must decline, etc.) were excluded from training exports.
  • Dominant Aegis 2.0 Resolution: For multi-label unsafe prompts in Nemotron, the annotator received the normalized ground-truth hint_violated_categories and independently selected the single most severe/dominant Aegis 2.0 category for unsafe_category, while violated_all preserves the full multi-label set.
  • Safe Polyfill for OpenAssistant & OpenThoughts: Because OpenAssistant/oasst1, OpenAssistant/oasst2, and open-thoughts/OpenThoughts3-1.2M do not contain unsafe violation flags, their safety columns are polyfilled as safe = True, unsafe_category = None, and violated_all = [] (synthetic = False for OpenAssistant; see the synthetic column definition for OpenThoughts).

3.1. OpenThoughts Pipeline (Classification → Translation → Cleaning)

  1. Candidate Pool & Level Filtering: A 60,000-prompt candidate pool was drawn from the code, math, and science shards of open-thoughts/OpenThoughts3-1.2M and interleaved round-robin (Code → Math → Science). Each candidate was classified with the Qwen3.5-9B rubric above, and only prompts scored level >= 3 were accepted until per-level quotas were met (4,247 accepted English prompts).
  2. Translation into 11 Languages: Each accepted English prompt was translated into ar, de, es, fr, hi, it, ja, ko, nl, th, zh-CN with tencent/Hy-MT2-7B-GGUF (Q4_K_M, temperature = 0.7, top_p = 0.6, top_k = 20, repetition_penalty = 1.05, max_tokens = 4096, context_per_slot = 8192).
  3. Label Propagation: Translations inherit the English source prompt's level, category, requires_reasoning, and (English) reason. Every source prompt therefore forms a 12-row translation group with identical labels.
  4. Group-Level Cleaning (4,247 → 3,311 groups, 50,964 → 39,732 rows): Entire 12-row groups were removed when the English source prompt was an exact duplicate of another source prompt (including duplicates with conflicting level labels), contained <image> placeholders (references to images unavailable in text-only prompts), or had any broken / truncated translation.

4. Dataset Construction: Safety, 12-Language & Complexity (level >= 3) Balancing

Raw OASST1 (33,170 rows) and OASST2 (48,051 rows) contain 81,221 safe prompts across 23 languages, but are heavily skewed toward English (39,172 rows) and Spanish (17,047 rows). Meanwhile, nvidia/Nemotron-Safety-Guard-Dataset-v3 supports 12 balanced languages (ar, de, en, es, fr, hi, it, ja, ko, nl, th, zh-CN). Furthermore, natural prompt distributions follow a steep Zipfian curve dominated by level = 1 queries.

To construct the balanced default (Route4Secure.parquet) split across all three axes (12 languages, safety, and enriched level >= 3 complexity):

  1. 12-Language Alignment: OASST1 and OASST2 language codes were normalized (zh $\rightarrow$ zh-CN) and filtered to the 12 languages supported by Nemotron (excluding 12,024 rows in non-Nemotron languages).
  2. 100% Preservation of level >= 3 in OpenAssistant (10,052 rows): Rather than losing rare high-complexity prompts during language capping, 100% of all level >= 3 prompts (10,052 rows) across all 12 in-scope languages in OASST1 + OASST2 were retained. In English (6,409 level >= 3 rows), only level >= 3 prompts were taken from OASST and the 1,409 excess above the 5,000 language cap was swapped 1-for-1 by removing 1,409 level = 1 English rows from Nemotron. In Spanish (2,277 level >= 3 rows), the remaining 2,723 slots to reach 5,000 were sampled proportionally from level = 0 and level = 2 (0 rows from level = 1).
  3. 50/50 Safe / Unsafe Supplement (9,377 rows): To balance the safe OASST prompts against the initial 79,983-row stratified Nemotron pass (33,184 safe + 46,799 unsafe), a deterministic zero-overlap supplement of 9,377 valid unsafe Nemotron rows was collected.
  4. High-Complexity (level >= 3) Nemotron Mining & 1-for-1 Swap (8,009 rows): From the remaining 362,030 unprocessed Nemotron rows, candidates were scored per language using structural complexity signals (language-normalized length, multi-line structure, code/math/step-by-step/technical markers) and evaluated in round-robin order with Qwen3.5-9B, retaining only prompts verified as level >= 3. This yielded 8,009 additional level >= 3 Nemotron rows (3,533 L3, 3,350 L4, 1,040 L5, 61 L6, 25 L7), which were swapped 1-for-1 per (language, safe) against level = 1 rows in Nemotron. This reduced level = 1 share from 48.56% down to 36.63% and more than doubled level >= 3 coverage from 10.15% (11,400 rows) to 23.69% (26,611 rows) while keeping total dataset size (112,338 rows), 12-language balance, and near-50/50 safety (50.94% Safe / 49.06% Unsafe) intact.
  5. Expert-Tier (level 5–7) OpenThoughts Expansion (17,712 rows): After step 4, the top three complexity classes were still severely under-represented (L5 2.00%, L6 0.35%, L7 0.17%; a 214× ratio between the largest and smallest class). From the cleaned open_thoughts shard, whole 12-language translation groups were sampled (seed 42): 500 groups (6,000 rows) each for L5 and L6, and all 476 available L7 groups (5,712 rows). Three groups whose prompts are pure chemical formulas / LaTeX (identical text across all 12 languages) were excluded. These rows are appended (no swap), with zero overlap against the existing rows.
    • L5 / L6 / L7 grew from 2,247 / 397 / 192 to 8,247 / 6,397 / 5,904 rows (6.34% / 4.92% / 4.54%).
    • The largest-to-smallest class ratio dropped from 214× to 7.0×, and level entropy rose from 2.29 to 2.68 bits (max 3.00).
    • Because all appended rows are safe, the global safety split moved from 50.94% / 49.06% to 57.62% / 42.38%.
    • OpenThoughts rows occupy id 112,338–130,049, ordered group by group (languages in alphabetical order within each group). The translation group key is (id - 112338) // 12 (in the open_thoughts config: id // 12). Use it for group-aware train/validation/test splits to avoid cross-lingual leakage.

5. Dataset Statistics (default — 130,050 Rows)

5.1. Breakdown by Source & Safety

Source Dataset safe synthetic Rows Share (%)
nvidia/Nemotron-Safety-Guard-Dataset-v3 False False 40,541 31.17%
nvidia/Nemotron-Safety-Guard-Dataset-v3 True False 25,391 19.52%
open-thoughts/OpenThoughts3-1.2M True True 16,236 12.48%
OpenAssistant/oasst2 True False 14,885 11.45%
nvidia/Nemotron-Safety-Guard-Dataset-v3 False True 14,571 11.20%
OpenAssistant/oasst1 True False 9,502 7.31%
nvidia/Nemotron-Safety-Guard-Dataset-v3 True True 7,448 5.73%
open-thoughts/OpenThoughts3-1.2M True False 1,476 1.13%
Total 74,938 Safe / 55,112 Unsafe 91,795 Non-synthetic / 38,255 Synthetic 130,050 100.00%

5.2. Language Distribution (12 Languages)

Language Code Safe (True) Unsafe (False) Total Rows Share (%)
English en 10,386 3,790 14,176 10.90%
Spanish es 9,228 4,399 13,627 10.48%
Chinese (Simplified) zh-CN 8,568 4,420 12,988 9.99%
German de 7,993 4,805 12,798 9.84%
French fr 6,802 4,691 11,493 8.84%
Thai th 5,522 5,018 10,540 8.10%
Dutch nl 4,122 5,353 9,475 7.29%
Arabic ar 4,312 5,024 9,336 7.18%
Japanese ja 4,762 4,387 9,149 7.03%
Italian it 4,785 4,359 9,144 7.03%
Hindi hi 4,293 4,382 8,675 6.67%
Korean ko 4,165 4,484 8,649 6.65%

5.3. Complexity Level Distribution (level 0–7)

Level Tier Safe (True) Unsafe (False) Total Rows Share (%)
"0" Trivial 10,744 6,594 17,338 13.33%
"1" Trivial 9,725 31,429 41,154 31.64%
"2" Moderate 15,289 11,946 27,235 20.94%
"3" Moderate 9,027 1,961 10,988 8.45%
"4" Complex 10,140 2,647 12,787 9.83%
"5" Complex 7,896 351 8,247 6.34%
"6" Expert / Advanced 6,330 67 6,397 4.92%
"7" Expert / Advanced 5,787 117 5,904 4.54%
  • level >= 3: 44,323 (34.08%) · level >= 5: 20,548 (15.80%)
  • requires_reasoning = False: 96,747 (74.39%)
  • requires_reasoning = True: 33,303 (25.61%)

5.4. Task Intent Category Distribution (category)

Category Rows Share (%)
Other 32,441 24.95%
Information & Science 24,492 18.83%
Writing & Content Creation 20,950 16.11%
Security & Hijack 14,399 11.07%
Coding & Software 11,292 8.68%
Roleplaying 8,340 6.41%
Mathematics 6,982 5.37%
Logical Reasoning & Problem Solving 4,657 3.58%
Text Processing & Summarization 2,252 1.73%
Arts & Culture 1,955 1.50%
Data Manipulation 1,222 0.94%
Language Translation 682 0.52%
Media Generation 386 0.30%

5.5. Dominant Aegis 2.0 Safety Category Distribution (unsafe_category)

Aegis 2.0 Category Rows Share (%)
null (Safe / Non-violating) 74,837 57.54%
Criminal Planning/Confessions 16,449 12.65%
Hate/Identity Hate 7,712 5.93%
Harassment 4,119 3.17%
PII/Privacy 3,641 2.80%
Controlled/Regulated Substances 3,308 2.54%
Sexual 3,090 2.38%
Violence 2,790 2.15%
Suicide and Self Harm 2,297 1.77%
Profanity 1,821 1.40%
Guns/Illegal Weapons 1,723 1.32%
Unauthorized Advice 1,510 1.16%
Illegal Activity 1,455 1.12%
Immoral/Unethical 1,286 0.99%
Sexual Minor 805 0.62%
Fraud/Deception 791 0.61%
Political/Misinformation/Conspiracy 720 0.55%
Malware 537 0.41%
Manipulation 517 0.40%
Threat 314 0.24%
High Risk Gov. Decision Making 180 0.14%
Copyright/Trademark/Plagiarism 148 0.11%

6. Usage

Loading with Hugging Face datasets 

from datasets import load_dataset

# 1. Load the main Route4Secure dataset (130,050 rows)
ds = load_dataset("Squezilia/Route4Secure", split="train")
print(ds)

# 2. Load individual normalized source shards if needed
oasst1_ds = load_dataset("Squezilia/Route4Secure", "oasst1", split="train")              # 33,170 rows
oasst2_ds = load_dataset("Squezilia/Route4Secure", "oasst2", split="train")              # 48,051 rows
nemotron_ds = load_dataset("Squezilia/Route4Secure", "nemotron", split="train")          # 97,369 rows
open_thoughts_ds = load_dataset("Squezilia/Route4Secure", "open_thoughts", split="train")  # 39,732 rows

Reading Dictionary-Encoded Parquet Directly with pyarrow / pandas 

import pyarrow.parquet as pq

table = pq.read_table("hf://datasets/Squezilia/Route4Secure/Route4Secure.parquet")
df = table.to_pandas()
print(df[["prompt", "level", "category", "requires_reasoning", "safe", "unsafe_category", "language"]].head())

7. Limitations & Ethical Considerations

  • Adversarial & Sensitive Content Warning: Because roughly 42% of Route4Secure consists of unsafe and jailbreak prompts sourced from nvidia/Nemotron-Safety-Guard-Dataset-v3, the dataset contains explicit examples of hate speech, criminal planning, self-harm, violence, malware requests, and other harmful categories. It is intended strictly for training and evaluating safety guardrails, moderation classifiers, and LLM routers.
  • LLM Annotation Disagreements: Complexity (level), domain (category), requires_reasoning, and dominant unsafe_category annotations are generated by Qwen3.5-9B (Q4_K_M). While safe and violated_all preserve upstream ground-truth labels, unsafe_category reflects the annotator model's prediction and can occasionally disagree with the binary safe flag on borderline prompts (~6% disagreement rate in raw Nemotron). Users training strict binary safety classifiers should rely on safe (or len(violated_all) > 0) as the ground-truth safety target.
  • High-Complexity Safety & Domain Skew: 97.4% of level 5–7 rows are safe, and ~86% of them come from OpenThoughts (code, math, and science only). A naive classifier may learn shortcuts such as "expert-level ⇒ math/code ⇒ safe" or "long prompt ⇒ high level" (OpenThoughts L6/L7 prompts are much longer than other L6/L7 rows). High-complexity unsafe prompts (e.g., sophisticated multi-step jailbreaks) remain scarce; consider class- or source-aware reweighting and evaluate on held-out complex unsafe prompts.
  • Machine-Translated Content: 11/12 of every OpenThoughts group is a Hy-MT2-7B translation (synthetic = True) that inherits the English labels and English reason. Despite group-level cleaning, translations may contain residual errors, and math/code-heavy prompts naturally keep a large share of Latin script, LaTeX, and code in non-Latin-script languages.
  • Long Prompts: OpenThoughts prompts are long (median ~370 XLM-R tokens; ~37% exceed 512 tokens). Encoder models with a 512-token limit will truncate them; head+tail truncation is recommended so that the final question is preserved.
  • Residual Language Ratio Skew: The 12 languages each represent 6.6%–10.9% of the dataset, but languages with heavy OpenAssistant coverage (en, es, zh-CN, de) have a higher internal proportion of safe prompts (~62%–73% safe) compared to languages sourced primarily from Nemotron (nl, ar, ko, hi, ~44%–50% safe).

8. License & Attribution

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license, consistent with upstream attribution requirements:

Citation 

@dataset{squezilia_route4secure_2026,
  author       = {Squezilia},
  title        = {Route4Secure: Multilingual Prompt Complexity, Intent and Safety Routing Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Squezilia/Route4Secure}}
}