nuhmanpk/semantic-router-dataset

Semantic Router Dataset for Tool Calling LLMs

A high-quality synthetic routing dataset designed for fine-tuning small language models such as Qwen2.5-0.5B, Qwen2.5-1.5B, Gemma, TinyLlama, and other instruction-tuned models for semantic tool routing.

This dataset trains models to select the correct tool from a dynamically changing tool pool given a natural language user query.

---

What This Dataset Is

Modern AI agents rely on routers to decide:

• which tool to call
• whether external actions are needed
• which API/function best matches user intent

This dataset focuses specifically on:

• semantic intent routing
• tool selection
• instruction-following
• ambiguity handling
• small-model alignment

The dataset is formatted for:

• Unsloth
• Hugging Face Transformers
• TRL / SFTTrainer
• Qwen ChatML
• LoRA fine-tuning

---

Task Format

The model receives:

• a system prompt containing available tools
• a user request
• and must output ONLY JSON

Example:

{
  "tool": "task_manager",
  "confidence": 0.91
}

---

Example Sample

Input

USER:
What's my highest priority task right now?

Expected Output

{
  "tool": "task_manager",
  "confidence": 0.91
}

---

Dataset Structure

The dataset uses ChatML conversation formatting:

{
  "conversations": [
    {
      "role": "system",
      "content": "tool definitions..."
    },
    {
      "role": "user",
      "content": "user request..."
    },
    {
      "role": "assistant",
      "content": "{\"tool\":\"task_manager\",\"confidence\":0.91}"
    }
  ]
}

Additional metadata:

• _correct_tool
• _category
• _source
• _pool

---

Included Tool Categories

The dataset currently contains 36 tools across multiple domains:

Code

• code execution
• debugging
• git
• regex
• Docker
• dependency management

Search

• web search
• arXiv search
• documentation lookup
• Wikipedia lookup
• product search

Productivity

• calendar management
• meeting scheduling
• email drafting
• note taking
• task management

Data

• SQL generation
• CSV analysis
• visualization
• statistics

Files

• PDF extraction
• image analysis
• OCR
• transcription

Communication

• Slack messaging
• GitHub issues
• Jira tickets

Math

• calculator
• equation solving
• unit conversion

General

• summarization
• translation
• sentiment analysis
• creative writing
• QA

---

Features

Dynamic Tool Pools

Each example contains a randomized tool pool.

This prevents the model from memorizing fixed label ordering and improves generalization.

---

Hard Negatives

The dataset includes intentionally ambiguous routing examples:

• calculator vs equation_solver
• web_search vs wikipedia_lookup
• task_manager vs note_taker
• sql_generator vs csv_analyzer

This significantly improves routing robustness.

---

Small-Model Friendly

Designed specifically for:

• 0.5B–3B models
• LoRA fine-tuning
• low VRAM training
• quantized inference

---

Use Cases

This dataset can be used for:

• AI agents
• semantic routers
• function calling
• MCP routing
• API selection
• tool-aware assistants
• autonomous workflows
• multi-tool orchestration
• retrieval routing
• lightweight local assistants

---

Training With Unsloth

Example:

from unsloth import FastLanguageModel
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-0.5B-Instruct",
    max_seq_length=2048,
    load_in_4bit=True,
)

dataset = load_dataset("nuhmanpk/semantic-router-dataset")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    args=SFTConfig(
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        num_train_epochs=3,
        learning_rate=2e-4,
        output_dir="./router_checkpoints",
    ),
)

trainer.train()

---

Loading the Dataset

Hugging Face

from datasets import load_dataset

dataset = load_dataset("nuhmanpk/semantic-router-dataset")

Local Disk

from datasets import load_from_disk

dataset = load_from_disk("./router_dataset/hf_dataset")

---

Dataset Splits

• Train
• Validation
• Test

The splits are stratified across tools to preserve routing diversity.

---

Generation Pipeline

The dataset was generated using:

• template-based synthesis
• semantic hard negatives
• randomized tool pools
• synthetic ambiguity injection
• quality filtering
• deduplication
• ChatML formatting

Optional augmentation can be performed with:

• Qwen
• Gemma
• Mistral
• TinyLlama
• other open instruction models

---

Recommended Models

Works especially well with:

• Qwen2.5-0.5B
• Qwen2.5-1.5B
• Gemma
• TinyLlama
• Phi models
• Llama 3.2 small variants

---

Future Improvements

Planned future versions may include:

• multi-tool routing
• clarification_required labels
• direct_response labels
• multilingual queries
• typo/noisy user input
• conversational follow-ups
• retrieval-based routing
• ranking-style supervision

---

License

Intended for research and educational purposes.

Please verify outputs before production deployment.

---

Author

Created by Nuhman PK

Focused on lightweight AI systems, routing architectures, and small-model tooling.