vllm-sr/router-signal-suite

Router signal suite

Training and test data for the ten vLLM Semantic Router signals: domain, jailbreak, safety, hazard, fact check, modality, PII, feedback, hallucination and tool need. The label cannot be read off the corpus a row came from, and every signal has held-out corpora. The design is in semantic-router#4305, and the comparison scored on it is in semantic-router#4306.

The builder generates no rows. Every row comes from a public corpus at a pinned revision. Text is published here only where the source's licence and access terms allow it (see Licences). Everything else is rebuilt from the sources with the builder in this repository, and the manifests check the rebuild row for row.

Contents

path	what it is
builder/	the builder, fetch_locked.py, build_all.sh and the lock files
builder/env/	Dockerfile and requirements.lock of the environment the published files were built in
manifests/suite/<task>/<file>.parquet	one row per suite row: id, task, source, subsource, label or labels, script, lang, tokens, length bin, group_sha256, text_sha256
manifests/fresh/, manifests/additions/	the same for the fresh held-out set and the jailbreak and PII additions
manifests/*/FILES.sha256	sha256 of every built JSONL file
text/<source>/<task>/<file>.jsonl	the rows of every source whose terms allow redistribution, each line as built
sources.json	every source: repositories and revisions, licence, gating, rows per file, and whether its text is published
predictions/, scoring/, tables.sh	per-row scores of every model in #4306's tables and the scripts that rebuild those tables
tables/	the output of one tables.sh run on a rebuilt suite
latency/	the latency and serving measurements behind #4306's latency table
tools/manifest.py	writes and verifies the manifests

text_sha256 is the sha256 of the canonical JSON (sorted keys, compact separators, UTF-8) of a row's text fields: text, text_pair, question, tools and spans. group_sha256 is the sha256 of the row's group. Rows that share a group stay on one side of every split.

The suite

10 tasks, 95 files, 536,508 rows, of which 398,156 are training rows.

task	files	rows
domain	10	104,802
fact_check	8	13,661
feedback	5	42,528
hallucination	10	37,905
hazard	9	51,221
jailbreak	10	25,621
modality	12	79,102
pii	10	60,189
safety	15	94,825
tool_need	6	26,654

Each task directory holds:

• train.jsonl for training.
• dev.jsonl for checkpoint selection, calibration and thresholds only.
• test.jsonl, the in-distribution test.
• test-<corpus>.jsonl, a training corpus's own published test split.
• hold-<corpus>.jsonl, a corpus, or a held-out slice of a training corpus, that no model trains on.

The file decides a row's role. Each row also carries a split field, which is the group-hash split before the dev and test caps. Rows past those caps go back to train and keep that field, so a row in train.jsonl can say test.

Row fields:

• id, task, text.
• label, or labels for hazard, which is multi-label.
• source, subsource, group, script, lang, tokens, length_bin, split.
• Hallucination rows add question and text_pair (the answer or claim). Tool-need rows add tools. PII rows add spans, character offsets into text.

The jailbreak and PII additions

vllm-sr/Decision-1.0-Route-0.6B trained on the suite's training files with two of them changed:

• Jailbreak train.jsonl adds 2,842 rows from the SaTML 2024 CTF and prompts.chat, which makes 16,368 rows.
• PII train.jsonl adds 3,804 rows from Few-NERD, which makes 47,736 rows.

Every other file is the suite's. The added rows were checked against every suite row of every task and split, and exact and near duplicates (MinHash, Jaccard 0.8) were dropped.

The fresh held-out set

44 files and 57,805 rows from public corpora that the suite never touched. Exact and near duplicates of any suite row were removed. The set was frozen before any further training and scored once (#4306). It is for evaluation only.

Rebuilding

The build runs on CPU and downloads 36.2 GB: only the files the builders read, at the revisions the lock files pin (builder/sources.lock.json for the suite, builder/fresh.lock.json and builder/additions.lock.json). Four sources are gated, so the token you pass must belong to an account that has accepted their terms: allenai/wildguardmix, allenai/wildjailbreak, hackaprompt/hackaprompt-dataset and, for the fresh held-out set, lytang/LLM-AggreFact. The llm-semantic-router/* repositories in the lock now redirect to vllm-sr/*. They keep their old names here because the download directories are named after them.

docker build -t suite-builder builder/env
mkdir -p raw out
docker run --rm --user "$(id -u):$(id -g)" -e HOME=/tmp -e HF_TOKEN -v "$PWD:/w" -w /w suite-builder \
  python3 builder/fetch_locked.py --raw raw \
    --lock builder/sources.lock.json builder/fresh.lock.json builder/additions.lock.json
docker run --rm --network none --user "$(id -u):$(id -g)" -e HOME=/tmp -v "$PWD:/w" -w /w suite-builder \
  builder/build_all.sh raw out

fetch_locked.py downloads every source at its pinned revision. It exits non-zero and names every source it could not fetch and every file whose sha256 differs from the lock. build_all.sh checks the download again, then builds the suite, the fresh held-out set and the two additions offline. It ends by comparing every built file with manifests/, and names the first row that differs.

These commands were run on 2026-10-04 on a second machine (16 cores, 30 GB of memory) from an empty download directory. The fetch took 34 GiB and found every file's sha256 as locked. build_all.sh with one job took 45 minutes (the largest builder needs about 10 GB of memory). All 95 suite files, all 44 fresh files and both addition files came out byte-identical to the published manifests.

The tables in #4306 and on the model card

tables.sh rebuilds them from a built suite, the fresh held-out set and predictions/, on CPU in about 45 minutes:

docker run --rm --user "$(id -u):$(id -g)" -e HOME=/tmp -v "$PWD:/w" -w /w suite-builder \
  bash tables.sh out/suite out/fresh/tasks tables-out

output	table
main_table.md, from compare_main.json and compare_released_mmbert32k.json	#4306's main table
compare_fresh.json	#4306's follow-up table on the fresh held-out set, and the tallies in the card's evaluation/FRESH.md
compare_tuned.json	the card's tallies on the suite
summary_suite/, summary_fresh/	the card's per-signal means and their paired differences
card_results.md	the per-file table in the card's evaluation/RESULTS.md
tally_shield_auc.json, tally_shield_pauc5.json	the Vela Shield calls in #4306's partial-AUC paragraph

tables/ holds a run on the rebuilt suite. Its comparison files equal the ones #4306's tables were computed from, and every cell of the main table, the follow-up table and the card's evaluation tables matches. The pairs in each compare_adjusted.py call draw their bootstrap replicates from one generator in turn, so the pair order in tables.sh is part of the result.

Each prediction file holds one line per scored row, {"id": ..., "scores": {label: probability}}, under predictions/<suite|fresh>/<arm>/<task>/<file>.pred.jsonl. Dev predictions, which set thresholds, sit in <arm>-dev where an arm has its own. The Decision arms have no scores for 5 rows of hold-halubench and 9 rows of fresh-seahorse. Their dev files cover part of a task's dev split: 300 of 1,288 hazard rows, 1,500 of 1,506 PII rows, and for the first release 1,464 of 1,500 safety rows. A comparison scores only the rows both arms have. The arms:

arm	model
decision-train-split	the Decision model in #4306's main table, trained on the suite's training split
decision-one-mixed-run	the same rows in one mixed run
decision-first-release	the first release of the Decision model on Hugging Face, which also trained on the in-distribution tests
decision-first-release-served	the first release as semantic-router#4086 serves it, fresh set only
decision-tuned	the tuned model, the weights now in vllm-sr/Decision-1.0-Route-0.6B
vela-matched-budget, vela-documented-recipe, mmbert-32k, tfidf	encoders and the words-only control, trained on the same rows
released-vela, vela-shield, released-mmbert32k	the released router models, scored as published

Latency

latency/cpu/ and latency/gpu/ are the measurements behind #4306's latency table, taken on one host (Xeon Platinum 8568Y+ with 8 pinned threads, and an MI300X): 205 requests one at a time in fp32, the median of the 200 after warm-up. The Decision model's CPU numbers come from cpu/decision-perkind.json, its GPU numbers from gpu/decision.json, and the eight Vela checkpoints' from cpu/vela.json and gpu/vela.json. The batched serving rates come from gpu/serving-4086-executor.json, gpu/serving-4086-scheduler.json and gpu/serving-4310-scheduler.json. They depend on the hardware, so they can be checked against these files but not rebuilt from them.

Licences

The text of 117 of the 139 sources is in text/: 517,049 of 600,959 rows (86.0%). Each text/<source>/ directory stays under that source's licence, which sources.json gives with the repository and revision. A text line is the built file's line, so its text_sha256 checks against the manifests. train-additions.jsonl holds only the rows the jailbreak and PII additions add. Their base rows are the suite's train.jsonl.

Restrictions that travel with the text:

• Non-commercial, for non-commercial use only: Alpaca, APIGen-MT, ArabicMMLU, BeaverTails evaluation, C-Eval, Emu Edit, FaithBench, HaluBench, IndicSafe, IndoMMLU, LinguaSafe, No Robots, OneIG-Bench and ToxicChat (14 sources, 28,968 rows).
• Share-alike, so adaptations keep the same licence: BIPIA, Dolly, EXAMS, Few-NERD, MGSM, RefChecker, Schema-Guided Dialogue, SHROOM, Stack Exchange and UNER English-EWT. Stack Exchange posts are CC BY-SA. A row's group is <site>:<question_id>, and its post is https://<site>.stackexchange.com/q/<question_id>.

The other 22 sources are published as manifests only. Their rows rebuild from the sources:

• Gated: WildGuardMix and WildJailbreak (AI2 Responsible Use Guidelines), HackAPrompt, and LLM-AggreFact (evaluation only, CC BY-ND 4.0).
• No-derivatives licence: KMMLU.
• No licence stated, or unknown: AnyInstruct, ELI5-Category, Seahorse, the router's modality dataset, Yahoo Answers, DSTC2, AlignBench and Tensor Trust.
• Prompts from LMSYS-Chat-1M, whose agreement prohibits transfer: PolyGuardMix's in-the-wild part and FactBench.
• Bound to the Llama Community License: ai4privacy's 500k set.
• Research only: the BELLE evaluation set and Lakera b3.

Known limits

• Hub history can be squashed and GitHub commits can become unreachable. The per-file hashes make a changed source detectable, not recoverable.
• Four jailbreak held-out sets (NotInject, PromptShield's test split, JailbreakHub and ToxicChat) were looked at during development. The fresh held-out set exists for that reason.
• Many labels come from a model. The builder's cell check bounds their error per corpus and label only.
• CoCoNot's card lists two licences (AI2 ImpACT low-risk and ODC-BY).
• Two fact-check held-out files give their label away without the text. In hold-no_robots, length alone separates the classes (AUC 0.977, mean 64 against 804 characters), and in hold-wildbench the WildBench category decides the label (Information seeking is always FACT_CHECK_NEEDED). Scores on them, including the fact-check rows behind #4306's tables, partly measure that. See semantic-router#4301.