license: mit
task_categories:

• text-generation
  language:
• en
  tags:
• math
• reasoning
• chain-of-thought
• tool-use
• activation-steering
  size_categories:
• n<1K

Tool Call Steering Data

Analyzed subset of OpenMathReasoning for studying tool call behavior in mathematical reasoning.

Dataset Contents

Data Files

• data/analyzed_subset.parquet - 500 matched CoT/TIR solution pairs
  • Top 250: Cases where CoT is much longer than TIR (tool use helped)
  • Bottom 250: Cases where TIR is much longer than CoT (tool use didn't help)

Analysis Results

• results/cot_tir_analysis_top.json - Analysis of top 250 pairs
• results/cot_tir_analysis_bottom.json - Analysis of bottom 250 pairs
• results/cot_tir_analysis_combined.json - Combined results
• results/label_mapping.json - Standardized tool call labels

Usage

from datasets import load_dataset
import pandas as pd

# Load the analyzed subset
ds = load_dataset("ncrispino/tool-call-steering-data", data_files="data/analyzed_subset.parquet")

# Or load as pandas
df = pd.read_parquet("hf://datasets/ncrispino/tool-call-steering-data/data/analyzed_subset.parquet")

# Load analysis results
import json
from huggingface_hub import hf_hub_download

label_mapping_path = hf_hub_download(
    repo_id="ncrispino/tool-call-steering-data",
    filename="results/label_mapping.json",
    repo_type="dataset"
)
with open(label_mapping_path) as f:
    label_mapping = json.load(f)

Tool Call Labels

Each tool call is labeled with its value:

Source

Derived from NVIDIA OpenMathReasoning.

License

MIT