ai-mitra/llm-router-dataset

llm-router dataset

Training data for ai-mitra/llm-router, a prompt task-classifier used by the llm-router Python package to route prompts to the best-fit LLM in agentic AI systems.

Each row is a prompt labeled with the task category it belongs to.

Labels

simple, coding, reasoning, security, summarization

Files

File	Rows	Purpose
training_data.jsonl	1700 (340/label)	Used to train the classifier: 35 hand-written examples per category plus ~305 template-generated ones per category.
regression_dataset.jsonl	25 (5/label)	Held out from training entirely -- hand-written, used only to sanity-check releases and calibrate confidence thresholds.

Format (one JSON object per line):

{"text": "Explain what a Python decorator does", "label": "coding"}

Loading

from datasets import load_dataset

ds = load_dataset(
    "ai-mitra/llm-router-dataset",
    data_files={"train": "training_data.jsonl", "regression": "regression_dataset.jsonl"},
)
print(ds["train"][0])

How it was built

training_data.jsonl combines:

1. 35 hand-written examples per category -- see build_dataset.py in the source repository's training/ directory.
2. ~305 template-generated examples per category -- combinatorial expansion of hand-written templates x filler vocabularies (languages, frameworks, countries, vulnerability types, document types, etc.), see training/augment_dataset.py. Generation is deterministic (fixed random seed) and deduplicates against both the hand-written set and regression_dataset.jsonl, so the regression set never leaks into training.

Validated before release for: empty examples, duplicate examples, missing labels, unknown labels, class imbalance, and train/regression leakage (see training/dataset_validation.py in the source repo).

Caveats

This is a synthetic dataset, not sourced from real user traffic. Template generation gives broad lexical coverage per category but not the full diversity of real-world phrasing -- expect the classifier's confidence to be more conservative on prompts phrased very differently from anything here. Extend training/augment_dataset.py with more templates/fillers (or add hand-written examples) to grow coverage.

License

MIT