SolvrenAI/class1-router-synthetic

Class 1 Decider router dataset (synthetic)

This dataset holds request-routing examples for Solvren AI's "Class 1 Decider", a lightweight router that sits in front of LLM calls and decides which path handles each request. Each row is one user request labelled with one of four routes:

Route	Meaning
small-model	simple, well-scoped, low-risk text work (reformat, short summary, extract, translate, classify)
frontier-model	complex multi-step reasoning, long-form drafting, nuanced analysis
retrieval	needs specific internal documents, records, or current external sources
refuse-or-escalate	high-stakes individualized advice, another person's private records, or anything that needs a qualified human

Splits

Config	Split	File	Rows	How it was made
default	train	train.jsonl	1,024	templates
default	validation	valid.jsonl	256	templates, unseen core requests
default	test	test.jsonl	256	templates, unseen core requests
handwritten	test	handwritten_test.jsonl	50	written by hand, different style

The three template splits are balanced: 25% of rows per route. The handwritten test split has 12 or 13 rows per route. It is a separate config because its rows carry an extra source column.

from datasets import load_dataset

ds = load_dataset("SolvrenAI/class1-router-synthetic")                                # train / validation / test
hw = load_dataset("SolvrenAI/class1-router-synthetic", "handwritten", split="test")  # 50 hand-written rows

How it was built

• Template splits: generated offline by local/gen_router_data.py, with no model calls. The grid is 4 domains (healthcare, legal, fintech, biotech) x 4 routes x 12 core requests, and each core request gets 8 phrasing variants (a greeting, an urgency marker, a team prefix and so on) plus fictional slot values (drug name, contract type, assay, record id).
• Leakage control: splitting is by core request. A request and all of its phrasings land in exactly one split, with 8 core requests per domain and route in train, 2 in validation and 2 in test, so the test requests never appear in training in any phrasing.
• handwritten config: 50 requests written by hand to check generalisation beyond the template family. They have different phrasing, varied length, typos and lowercase, and a mix of domains that includes general business requests outside the four training domains. None of them appears in train.

Format

Each row uses the chat messages format (as in the upstream aws-samples platform and mlx-lm):

{"messages": [
  {"role": "system", "content": "<routing policy prompt>"},
  {"role": "user", "content": "Extract the follow-up date from this lab report."},
  {"role": "assistant", "content": "small-model"}],
 "domain": "healthcare"}

Every row carries the same system prompt, which defines the four routes. Rows in the handwritten config also carry "source": "handwritten".

Privacy and safety

• No real PII or PHI. Every name, record id, company, lot number and amount is a fictional placeholder. The company names (Acme, Northwind, Contoso, Globex, Initech) are the usual sample names.
• refuse-or-escalate examples are benign escalation triggers: urgent symptoms, a request for an individualized medical, legal or financial decision, a request for another person's records, a skipped human approval, or someone who needs a person rather than a model. The dataset contains no harmful instructions.

Limitations

• Template splits are easy. A model can score near 100% on test by learning the template family, so read the handwritten config for a more honest signal.
• English only. The labels reflect one opinionated routing policy, and some requests sit near a boundary (for example, a short drafting task could reasonably be small-model or frontier-model).
• Not a safety benchmark. Do not use it to evaluate content-safety classifiers.

License

MIT-0, the same as the source repository.

Citation

If you use this dataset in your work, please cite:

@misc{solvren2026class1routerdata,
  title        = {Class 1 Decider Router Synthetic Dataset: Request Routing for Regulated Domains},
  author       = {{Solvren AI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SolvrenAI/class1-router-synthetic}},
  note         = {Code: https://github.com/solvren-ai/class1-router}
}

Adapted from aws-samples/sample-slm-finetuning-and-evaluation-platform (MIT-0).