| --- |
| license: mit |
| language: |
| - en |
| pretty_name: To Call or Not to Call — Tool-Calling Evaluation Datasets |
| tags: |
| - tool-calling |
| - agents |
| - llm-evaluation |
| - function-calling |
| - reasoning |
| task_categories: |
| - question-answering |
| - text-generation |
| configs: |
| - config_name: entity |
| data_files: |
| - split: train |
| path: entity/entity.jsonl |
| - config_name: bfcl |
| data_files: |
| - split: train |
| path: bfcl/bfcl.jsonl |
| - config_name: synthetic_multiplication |
| data_files: |
| - split: train |
| path: synthetic_multiplication/synthetic_multiplication.jsonl |
| - config_name: synthetic_nn |
| data_files: |
| - split: train |
| path: synthetic_nn/synthetic_nn.jsonl |
| --- |
| |
| # To Call or Not to Call: Tool-Calling Evaluation Datasets |
|
|
| Datasets accompanying the paper **["To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling"](https://arxiv.org/abs/2605.00737)** (Wu et al., 2026). |
|
|
| The paper evaluates when language models should call external tools (web search, calculator) rather than answering from parametric knowledge, from both a **normative** perspective (when is a tool call truly needed/useful?) and a **descriptive** perspective (when does the model *think* it needs a tool, and how well does that judgment track the truth?). |
|
|
| Code: https://github.com/QinyuanWu0710/ToCall_or_NotToCall |
|
|
| This repository hosts the four benchmark inputs built or curated for that framework. It does **not** include the InVivo query set referenced in the code repository — that data is derived from real user conversations and is withheld pending an internal PII review. |
|
|
| It also does not include the GSM-Hard calculator benchmark used in the paper's calculator experiments, since that dataset is third-party data already published at [reasoning-machines/gsm-hard](https://huggingface.co/datasets/reasoning-machines/gsm-hard); the code repository downloads it directly from there. |
|
|
| ## Configs / Subsets |
|
|
| Load any subset with: |
|
|
| ```python |
| from datasets import load_dataset |
| |
| entity = load_dataset("QinyuanWu0710/ToCall_or_NotToCall", "entity") |
| bfcl = load_dataset("QinyuanWu0710/ToCall_or_NotToCall", "bfcl") |
| synth_mult = load_dataset("QinyuanWu0710/ToCall_or_NotToCall", "synthetic_multiplication") |
| synth_nn = load_dataset("QinyuanWu0710/ToCall_or_NotToCall", "synthetic_nn") |
| ``` |
|
|
| ### `entity` (500 rows) |
|
|
| Entity-centric knowledge queries used for the web-search experiments (no-search / auto-search / force-search conditions). Entities were extracted from real-world conversational queries, classified into topical categories, and de-identified down to the entity mention itself — no conversation text or user-identifying information is included. |
|
|
| | Column | Type | Description | |
| |---|---|---| |
| | `entity_text` | string | The entity mention (e.g. a person, organization, place, or term) that seeds the query. Combined with `ENTITY_QUERY_PROMPT_TEMPLATE` in the code repo to form the full user question. | |
| | `category` | string | Topical category assigned to the entity (e.g. *People & Personal Attributes*, *Finance-related*, *Culture & Entertainment*). | |
| | `count` | int | Number of times this entity was observed during dataset construction (a rough popularity signal). | |
|
|
| ### `bfcl` (314 rows) |
|
|
| Chained factual-lookup questions in the style of the Berkeley Function-Calling Leaderboard: each question requires looking up one fact, then using that fact to answer the next question in the chain (e.g. *"Who won Best Picture at the 2024 Academy Awards?" → "Who is the director of \[that film\]?"*). Designed so a model without live web access will systematically drift from the correct chain as questions get further from its training cutoff. |
|
|
| | Column | Type | Description | |
| |---|---|---| |
| | `question` | string | The question to answer. | |
| | `correct_response` | string | Ground-truth answer used for exact/fuzzy-match scoring. | |
|
|
| ### `synthetic_multiplication` (1,000 rows) |
| |
| Synthetic multi-factor multiplication problems (GSM-style final-answer format) used for the calculator tool-use experiments, generated by `data/generate_multiplication_dataset.py` in the code repo. Difficulty scales with the number of factors and their digit lengths, giving a controllable necessity signal for "does this task actually require a calculator." |
| |
| | Column | Type | Description | |
| |---|---|---| |
| | `question` | string | Prompt containing the multiplication expression, e.g. `"Compute the product: 408 * 872 * 901"`. | |
| | `answer` | string | Ground-truth answer in GSM final-answer format, e.g. `"#### 46497576972551424"`. | |
| | `difficulty` | string | `easy`, `medium`, or `hard`. | |
| | `num_multiplications` | int | Number of multiplication operations in the expression. | |
| | `operand_digits` | list[int] | Digit length of each operand, in order. | |
| | `expression` | string | The raw arithmetic expression, without the surrounding prompt text. | |
|
|
| ### `synthetic_nn` (1,000 rows) |
| |
| Synthetic large-integer squaring problems (`N * N` for large `N`), the harder companion set to `synthetic_multiplication`. Difficulty is labeled by operand digit length (e.g. `27-digit`) rather than by named tier, since a single repeated-operand square already saturates non-tool arithmetic well before 1,000-digit multiplication does. |
|
|
| | Column | Type | Description | |
| |---|---|---| |
| | `question` | string | Prompt containing the squaring expression, e.g. `"Compute the product: 644969583704402664902690654 * 644969583704402664902690654"`. | |
| | `answer` | string | Ground-truth answer in GSM final-answer format. | |
| | `difficulty` | string | Digit-length label of the operand, e.g. `"27-digit"`. | |
| | `num_multiplications` | int | Always `1` for this subset (a single squaring operation). | |
| | `operand_digits` | list[int] | Digit length of each of the two (identical) operands. | |
| | `expression` | string | The raw arithmetic expression, without the surrounding prompt text. | |
|
|
| Both synthetic sets are used with `--task synthetic_multiplication` in the code repo; which file is scored is controlled by pointing `--data`/`DATA` at the corresponding JSONL file. |
|
|
| ## Intended Use |
|
|
| These datasets are intended for evaluating and studying **when an LLM agent should invoke external tools** (web search, calculator) versus answering from parametric knowledge — e.g. reproducing the paper's necessity/utility/affordability analysis, or as inputs to other tool-use / agentic-evaluation research. `entity` and `bfcl` are best paired with a web-search tool; `synthetic_multiplication` and `synthetic_nn` are best paired with a calculator tool. |
|
|
| ## License |
|
|
| Released under the MIT License, matching the code repository. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{wu2026call, |
| title={To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling}, |
| author={Wu, Qinyuan and Das, Soumi and Amani, Mahsa and Nag, Arijit and Lee, Seungeon and Gummadi, Krishna P and Ravichander, Abhilasha and Zafar, Muhammad Bilal}, |
| journal={arXiv preprint arXiv:2605.00737}, |
| year={2026} |
| } |
| ``` |
|
|
