alucent/mirror-ToolRM-train-data

ToolRM Training Dataset

📖 Dataset Description

This is a version of the training data utilized for ToolRM, a collection of outcome reward models specifically designed for evaluating and improving function-calling capabilities in large language models. It consists of ~459K examples, where each example includes a user-assistant conversation, available tool specifications, and a pair of correct and incorrect tool calls. The incorrect calls were generated by prompting 9 open-source language models on queries from three public datasets. Reward Models trained on this dataset were found to result in an average improvement of up to 25% in downstream task performance, enhance robustness to input noise, and enable data-efficient fine-tuning through reward-guided filtering.

📊 Dataset Statistics

• Total Training Samples: 458,575
• Composition:
  • Single-turn interactions: 256,851 samples
  • Multi-turn interactions: 159,757 samples
  • Irrelevance cases: 41,967 samples
• Source Datasets: APIGen, Schema-Guided Dialogue (SGD), xlam-irrelevance
• Generator Models: 9 permissively-licensed open-weight models

🗂️ Dataset Schema

The dataset contains the following fields:

Field	Type	Description
uuid	str	Unique identifier for each training sample
dataset_name	str	Source dataset from which the sample was derived
conversation	list	Conversation between user and assistant
tools	str	Catalog of available function specifications
tool_calls_correct	str	Ground-truth correct tool invocations for the given conversation
tool_calls_incorrect	str	Incorrect tool invocations generated by the model pool
generator_model	str	Identifier of the model that produced the incorrect tool call

Note: tools, tool_calls_correct, and tool_calls_incorrect fields have been serialized. While loading the dataset, convert them into JSON objects using json.loads

⚙️ Data Generation Methodology

Generator Model Pool

The incorrect tool calls were generated using the following models:

• Granite Series: granite-3.3-2b-instruct, granite-3.3-8b-instruct, granite-20b-functioncalling
• SmolLM: SmolLM2-1.7B-Instruct, SmolLM3-3B
• Mistral Series: Mistral-7B-Instruct-v0.3, Mistral-Nemo-Instruct-2407
• GPT-OSS Series: gpt-oss-20b, gpt-oss-120b

Data Collection Process

1. Source Datasets: We start with publicly available function-calling datasets that cover a wide range of interaction patterns
2. Obfuscation: Function and parameter names were replaced with random strings, and schema keys were reordered to prevent models from regurgitating the training data
3. Generation: Each sample is processed through the model pool to generate function calls
4. Verification: The generated outputs are compared against ground-truth annotations to identify incorrect calls
5. Filtering: We keep only the incorrect generations, selecting up to three incorrect samples per query to maintain diversity while avoiding over-representation

🎯 Benchmark

In a Best-of-N setting, we found that ToolRM significantly improves performance over Greedy decoding, Majority Voting, and Schema Validation baselines.

For reward-guided data filtering, we found that a model fine-tuned with 8K top-ranked samples by ToolRM outperforms the model fine-tuned with the entire training dataset of 16K samples.

More experiments and a detailed discussion of the results can be found in the paper.

📚 Citation

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

@misc{agarwal2025toolrmoutcomereward,
      title={ToolRM: Outcome Reward Models for Tool-Calling Large Language Models},
      author={Mayank Agarwal and Ibrahim Abdelaziz and Kinjal Basu and Merve Unuvar and Luis A. Lastras and Yara Rizk and Pavan Kapanipathi},
      year={2025},
      eprint={2509.11963},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.11963},
}