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) with adversarial, synthetic, and real-world safety examples from NVIDIA Nemotron Safety Guard Dataset v3, 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): Balanced 50.94% Safe / 49.06% Unsafe distribution in the primary default configuration.
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 balanced main dataset (default) or inspect any individual source shard:

Config Name	File	Rows	Size	Languages	Safe / Unsafe Split	Description
default	Route4Secure.parquet	112,338	42.6 MB	12 balanced languages	50.94% Safe (57,226) / 49.06% Unsafe (55,112)	Main training dataset. Combines language- and complexity-balanced OASST1 + OASST2 with nemotron.
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).

---

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, or nvidia/Nemotron-Safety-Guard-Dataset-v3).
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), False for human-curated annotations (OASST1, OASST2, and human-labeled Nemotron rows).

---

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).

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: Because OpenAssistant/oasst1 and OpenAssistant/oasst2 are human-curated conversational datasets without unsafe violation flags, their safety columns are polyfilled as safe = True, synthetic = False, unsafe_category = None, and violated_all = [].

Complexity Rubric (level 0–7)

• 0–1 (Trivial): Simple greetings, 1-step direct factual questions, basic definitions.
• 2–3 (Moderate): Middle/high-school level math or riddles, basic code snippets, standard information lookup, straightforward writing/summarization.
• 4–5 (Complex): Multi-step reasoning, intermediate coding/debugging, undergraduate-level domain concepts, structured multi-constraint analysis.
• 6–7 (Expert / Advanced): Highly specialized research/domain expertise, complex multi-constraint algorithmic design, advanced theoretical math/science proofs, intricate systems architecture.

---

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, ~50/50 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. Dataset Statistics (default — 112,338 Rows)

5.1. Breakdown by Source & Safety

Source Dataset	safe	synthetic	Rows	Share (%)
nvidia/Nemotron-Safety-Guard-Dataset-v3	False	False	40,541	36.09%
nvidia/Nemotron-Safety-Guard-Dataset-v3	True	False	25,391	22.60%
OpenAssistant/oasst2	True	False	14,885	13.25%
nvidia/Nemotron-Safety-Guard-Dataset-v3	False	True	14,571	12.97%
OpenAssistant/oasst1	True	False	9,502	8.46%
nvidia/Nemotron-Safety-Guard-Dataset-v3	True	True	7,448	6.63%
Total	57,226 Safe / 55,112 Unsafe	90,319 Human / 22,019 Synthetic	112,338	100.00%

5.2. Language Distribution (12 Languages)

Language	Code	Safe (True)	Unsafe (False)	Total Rows	Share (%)
English	en	8,910	3,790	12,700	11.31%
Spanish	es	7,752	4,399	12,151	10.82%
Chinese (Simplified)	zh-CN	7,092	4,420	11,512	10.25%
German	de	6,517	4,805	11,322	10.08%
French	fr	5,326	4,691	10,017	8.92%
Thai	th	4,046	5,018	9,064	8.07%
Dutch	nl	2,646	5,353	7,999	7.12%
Arabic	ar	2,836	5,024	7,860	7.00%
Japanese	ja	3,286	4,387	7,673	6.83%
Italian	it	3,309	4,359	7,668	6.83%
Hindi	hi	2,817	4,382	7,199	6.41%
Korean	ko	2,689	4,484	7,173	6.39%

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	15.43%
"1"	Trivial	9,725	31,429	41,154	36.63%
"2"	Moderate	15,289	11,946	27,235	24.24%
"3"	Moderate	9,027	1,961	10,988	9.78%
"4"	Complex	10,140	2,647	12,787	11.38%
"5"	Complex	1,896	351	2,247	2.00%
"6"	Expert / Advanced	330	67	397	0.35%
"7"	Expert / Advanced	75	117	192	0.17%

• requires_reasoning = False: 96,675 (86.06%)
• requires_reasoning = True: 15,663 (13.94%)

5.4. Task Intent Category Distribution (category)

Category	Rows	Share (%)
Other	32,441	28.88%
Writing & Content Creation	20,950	18.65%
Information & Science	18,624	16.58%
Security & Hijack	14,399	12.82%
Roleplaying	8,340	7.42%
Coding & Software	5,724	5.10%
Logical Reasoning & Problem Solving	4,441	3.95%
Text Processing & Summarization	2,252	2.00%
Arts & Culture	1,955	1.74%
Data Manipulation	1,222	1.09%
Mathematics	934	0.83%
Language Translation	682	0.61%
Media Generation	374	0.33%

5.5. Dominant Aegis 2.0 Safety Category Distribution (unsafe_category)

Aegis 2.0 Category	Rows	Share (%)
null (Safe / Non-violating)	57,125	50.85%
Criminal Planning/Confessions	16,449	14.64%
Hate/Identity Hate	7,712	6.86%
Harassment	4,119	3.67%
PII/Privacy	3,641	3.24%
Controlled/Regulated Substances	3,308	2.94%
Sexual	3,090	2.75%
Violence	2,790	2.48%
Suicide and Self Harm	2,297	2.04%
Profanity	1,821	1.62%
Guns/Illegal Weapons	1,723	1.53%
Unauthorized Advice	1,510	1.34%
Illegal Activity	1,455	1.30%
Immoral/Unethical	1,286	1.14%
Sexual Minor	805	0.72%
Fraud/Deception	791	0.70%
Political/Misinformation/Conspiracy	720	0.64%
Malware	537	0.48%
Manipulation	517	0.46%
Threat	314	0.28%
High Risk Gov. Decision Making	180	0.16%
Copyright/Trademark/Plagiarism	148	0.13%

---

6. Usage

Loading with Hugging Face datasets

from datasets import load_dataset

# 1. Load the main balanced Route4Secure dataset (112,338 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

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 50% 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.
• Residual Language Ratio Skew: While Route4Secure is globally balanced at 50/50 safe/unsafe and each of the 12 languages represents 6.4%–11.3% of the dataset, languages with heavy OpenAssistant coverage (en, es, zh-CN, de) have a slightly higher internal proportion of safe prompts (~58%–62% safe) compared to languages sourced primarily from Nemotron (nl, ar, hi, ko, ~33%–39% 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:

• nvidia/Nemotron-Safety-Guard-Dataset-v3 — Licensed under CC-BY-4.0.
• OpenAssistant/oasst1 & OpenAssistant/oasst2 — Licensed under Apache-2.0.

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}}
}