datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
glaive_toolcall_enBorrowed from: https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2
You can use it in LLaMA Factory by specifying dataset: glaive_toolcall_en.
python-toolcallsLogs from run_python_code tool used for benchmarking.
Nexus-Agents-ToolCalling
Nexus Agents — Tool-Calling Conversations
Synthetic, schema-verified tool-calling conversations for training the Nexus Projects
agents. This is the exact data behind
Nemotron-3-Nano-30B-A3B — Nexus Agents (GGUF),
including the verification transcripts that scored it (27/27 on the behavioral
interview eval, vs 13/27 for the base model).
Links: the fine-tuned model →
Nemotron-3-Nano-30B-A3B — Nexus Agents (GGUF) ·
the generator + seed data + eval harness →
Nexus Training Studio ·… See the full description on the dataset page: https://huggingface.co/datasets/NexusProjectsAI/Nexus-Agents-ToolCalling.glaive_toolcall_enBorrowed from: https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2
You can use it in LLaMA Factory by specifying dataset: glaive_toolcall_en.
tool-calling-english-100k
Tool Calling English (100K)
100,000 tool-calling conversations in OpenAI function calling format — the largest general English tool-use dataset for fine-tuning.
Motivation
Models trained without tool-calling examples struggle in agentic deployments. This dataset trains the full cycle: deciding when to call a tool, calling it with correct arguments, interpreting the result, and producing a grounded final response.
Dataset Description
100,000… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/tool-calling-english-100k.glaive_toolcall_zhBorrowed from: https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2
Translated by GPT-3.5.
You can use it in LLaMA Factory by specifying dataset: glaive_toolcall_zh.
tool-calling-mix
This is a dataset for fine-tuning a language model to use tools. I combined sources from various other tool calling datasets and added some non-tool calling examples to prevent catastrophic forgetting.
Dataset Overview
Motivation
This dataset was created to address the need for a diverse, high-quality dataset for training language models in tool usage. By combining multiple sources and including non-tool examples, it aims to produce models that can effectively use tools… See the full description on the dataset page: https://huggingface.co/datasets/younissk/tool-calling-mix.pi-toolcall-dataset
Pi edit/write Tool-Call Training Dataset (Qwen3 / Unsloth Studio)
A conversational training dataset that teaches a Qwen3 ~35B base model to
behave as a pi coding agent: plan inside `
reasoning blocks, then issue correcteditandwritetool calls (plus supportingread/bash/grep/find/ls`), handle tool errors, and recover.
The system prompt, tool schemas, and exact tool result/error strings are
transcribed verbatim from the installed pi package
(@earendil-works/pi-coding-agent… See the full description on the dataset page: https://huggingface.co/datasets/marianbusoi/pi-toolcall-dataset.2026-07-31-toolcalling-tulu-20-80-mixture
Tool-calling + TULU3 replay SFT mixture (20/80) for Qwen3.6-27B
The training mixture behind
LASR-Callum/2026-07-31-wrongly-trained-qwen36-toolcalling-tulu-lora-20-80: 1,492,442 Qwen3.6
tokens across 2,002 pre-rendered conversations, split
19.96% agentic tool-use / 80.04% TULU3 replay.
Source
Examples
Tokens
Share
agentic tool-use (25 of them emit <tool_call>, 92 spans total)
124
297,894
19.96%
TULU3 replay
1,878
1,194,548
80.04%
Total
2,002
1,492,442… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-07-31-toolcalling-tulu-20-80-mixture.2026-07-31-toolcalling-tulu-sft-run
Run record — Qwen3.6-27B tool-calling 20/80 SFT
Everything the training run produced except the weights: the TRL log history, the resolved
config, the environment, the loss/accuracy figure and its greppable markdown mirror.
The adapter is at LASR-Callum/2026-07-31-wrongly-trained-qwen36-toolcalling-tulu-lora-20-80; the training data is
at LASR-Callum/2026-07-31-toolcalling-tulu-20-80-mixture.
Required metadata
field
value
experiment
One bf16 LoRA SFT… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-07-31-toolcalling-tulu-sft-run.model-toolcall-research
Model Toolcall Research
tool-call-efficiency
tool-call-efficiency
Made with the whileai SDK · Collections: Efficiency, Start here: foundational post-training datasets
Teach an agent to make every tool call count.
An agent that calls a tool twice with the same arguments, looks up what
the user just told it, or keeps calling after the task is done is slow,
expensive, and harder to trust. Ask a base Qwen3-4B to work through
1,133 tool-using tasks across six agents and it does this a lot:
only 52% of its 6,681 rollouts finish… See the full description on the dataset page: https://huggingface.co/datasets/while-ai/tool-call-efficiency.autoscientist-toolcaller-dataset
AutoScientist Tool-Calling Dataset
A curated function-calling / tool-use dataset for the Adaption AutoScientist Challenge. Its
distinguishing feature is a large slice of hard negatives and reliability-focused cases — where the
correct behavior is not a plain tool call.
Adaptive Data quality (real): on the fixed set (c4923b7f…, graded on 1,000 of 2,440 rows
under the free-tier cap) the platform reported 7.0 → 8.1, +15.7%, grade C → B — now confirmed by a
completed, uncapped run… See the full description on the dataset page: https://huggingface.co/datasets/pandeyankit84/autoscientist-toolcaller-dataset.SFT_glaive_toolcall_en
Preparing Your Dataset
Once you’ve decided that fine-tuning is the best approach—after optimizing your prompt as much as possible and identifying remaining model issues—you’ll need to prepare training data. Start by creating a diverse set of example conversations that mirror those the model will handle during production.
Each example should follow this structure below, consisting of a list of messages. Each message must include a role, content, and an optional name. Make sure some… See the full description on the dataset page: https://huggingface.co/datasets/GreenNode/SFT_glaive_toolcall_en.toolcall-datatoti-cakery-toolcall
Toti Cakery — Tool-Calling Fine-Tuning Dataset (Qwen3, v7)
Synthetic bilingual (Indonesian ~78% / English ~22%) SFT dataset for the Toti
Cakery WhatsApp chatbot: 13 LangChain tools (11 for customers, +2 owner-only
reports) and grounded answers from RAG FAQ context. Rows are built from the
live runtime code (SYSTEM_PROMPT, TOOL_REMINDER, tool schemas via
convert_to_openai_tool, _history_view, pertanyaan_dengan_konteks), so the
training prompt is byte-identical to what the model… See the full description on the dataset page: https://huggingface.co/datasets/LasagnaS/toti-cakery-toolcall.model-toolcall-research
Model Toolcall Research
This dataset stores newline-delimited agent traces from bounded research runs on model repository tool-schema support.
The Dataset Viewer is configured to index only .jsonl files:
toolcall_traces loads trace files under traces/**/*.jsonl.
research_session loads top-level provenance/session traces from *.jsonl.
The archive/ directory preserves the earlier .trace.json uploads for reference, but those files are newline-delimited JSON streams rather than… See the full description on the dataset page: https://huggingface.co/datasets/cfahlgren1/model-toolcall-research.toolcall_bench
When2Call
💾 Github | 📄 Paper
Dataset Description:
When2Call is a benchmark designed to evaluate tool-calling decision-making for large language models (LLMs), including when to generate a tool call, when to ask follow-up questions, when to admit the question can't be answered with the tools provided, and what to do if the question seems to require tool use but a tool call can't be made.
We find that state-of-the-art tool-calling LMs show significant room for… See the full description on the dataset page: https://huggingface.co/datasets/Compumacy/toolcall_bench.tech-duinn-toolcallsloap-reasoning-toolcalling-20k
loap-reasoning-toolcalling-20k
loap-reasoning-toolcalling-20k is a synthetic dataset designed to train language models in reasoning (Chain of Thought) and tool usage.
Language: English
Format: Chat (System, User, Model, Tool)
Dataset Structure
[
{
"id": "synthetic_agent_00001",
"conversations": [
{
"role": "system",
"content": "You are a helpful AI agent.\nYou have access to the following tools:"
},
{
"role": "tools"… See the full description on the dataset page: https://huggingface.co/datasets/igidn/loap-reasoning-toolcalling-20k.tool-calling-browser-agent-tasks
Dataset Card
Created by: DataCreator AI
Overview
Tool Calling for Agentic Tasks with Multi-Step Workflows contains 1,062 synthetic multi-turn conversations between a user and an AI assistant. The examples primarily focus on practical agentic tasks such as train ticket booking, dynamic form filling, and payment processing. It provides diverse scenarios including successful execution, context retrieval, tool integration, and failure recovery.
The dataset is… See the full description on the dataset page: https://huggingface.co/datasets/DataCreatorAI/tool-calling-browser-agent-tasks.scugnizz-toolcalling-synthetic-v3
Scugnizz Tool Calling Synthetic
Dataset sintetico per TOOL_CALL / TOOL_RESULT.
Categorie:
{
"negative_tool_not_available": 18,
"tool_result_mail": 2530,
"positive_hash": 6,
"similar_tools": 90,
"tool_result_finance": 164457,
"positive_ip": 15,
"tool_result_weather": 134612,
"positive_dns": 60,
"positive_multitool": 432,
"tool_result_calendar": 448,
"positive_weather": 72,
"negative_no_tool_needed": 3,
"negative_missing_required_arg": 3… See the full description on the dataset page: https://huggingface.co/datasets/ProjectScugnizz/scugnizz-toolcalling-synthetic-v3.tool-callingCodeXomics-ToolCalling-v1
CodeXomics-ToolCalling-v1
This dataset contains the supervised tool-calling trajectories used to fine-tune
CodeXomics-ToolAgent-4B-v1 (internally qwen3.5:4b-codexomics-tools-v5) for the CodeXomics
genomics workbench (an AI-native genome browser; source:
github.com/Scilence2022/CodeXomics). It is released
as the reproducibility artifact for the corresponding paper section and is licensed under
Apache-2.0.
Contents
train.jsonl — 373 supervised examples
valid.jsonl… See the full description on the dataset page: https://huggingface.co/datasets/CodeXomics/CodeXomics-ToolCalling-v1.tool-calling-traces-ptbr
Tool calling conversations in Portuguese
484 synthetic conversations that teach a model when to call a tool, which one to call and
with which arguments, and also when to answer directly, with no tool at all.
Each line of the file is a complete conversation: the user's question, the tool call, the
simulated return of that tool, and the final answer.
It was built because no dataset of tool calling in Portuguese with fictional tools existed.
The 30 tools and the user questions were… See the full description on the dataset page: https://huggingface.co/datasets/annajuliaasf/tool-calling-traces-ptbr.wish-engine-toolcall-hardening-v1
wish-engine-toolcall-hardening-v1
Targeted hardening dataset for wish-engine implementor tool-calling reliability.
Splits
train: 233
validation: 32
test: 29
Source mix
eval-mistake windows from prior benchmark failures
guardrail recovery trajectories (fix-guardrails segments)
generalized v3 tool-call examples
Generated by:
scripts/build-hf-toolcall-hardening-dataset.mjs
wish-engine-toolcall-id-selection-v1
wish-engine-toolcall-id-selection-v1
Targeted dataset for correcting canonical tool ID selection errors.
Splits
train: 167
validation: 16
test: 14
Source mix
v3 generalized tool-call rows selected by eval mistake IDs
canonical tool ID failures (hallucinated_unknown_tool)
analysis-only failures (missing_tool_call)
Generated by:
scripts/build-hf-toolcall-id-selection-dataset.mjs
wish-engine-toolcall-next-v3-strict-general
wish-engine-toolcall-next-v3-strict-general
Wish-engine implementor next-step tool-calling dataset (v3 strict generalization subset, dynamic aliases).
Splits
train.jsonl: 8081980 bytes
validation.jsonl: 1008543 bytes
test.jsonl: 997791 bytes
Schema
Rows are JSONL with at least:
id
messages (chat format with assistant tool_calls)
tool_name
metadata fields (mode, status, trajectory_*)
Notes
Tool names are dynamically aliased per sample.
A tool… See the full description on the dataset page: https://huggingface.co/datasets/sahilmob/wish-engine-toolcall-next-v3-strict-general.glaive-toolcall-mixwish-engine-toolcall-id-selection-v4-hardening
wish-engine-toolcall-id-selection-v4-hardening
Targeted dataset for correcting canonical tool-ID selection mistakes with synthetic hard negatives.
Splits
train: 322
validation: 33
test: 35
Source mix
v3 generalized tool-call rows selected by eval mistake IDs
ID-selection repair supervision from real benchmark failures
synthetic hard negatives with deterministic ID perturbation + roster reorder
Generated by:… See the full description on the dataset page: https://huggingface.co/datasets/sahilmob/wish-engine-toolcall-id-selection-v4-hardening.