datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
DeepScaleR-Preview-Dataset
Data
Our training dataset consists of approximately 40,000 unique mathematics problem-answer pairs compiled from:
AIME (American Invitational Mathematics Examination) problems (1984-2023)
AMC (American Mathematics Competition) problems (prior to 2023)
Omni-MATH dataset
Still dataset
Format
Each row in the JSON dataset contains:
problem: The mathematical question text, formatted with LaTeX notation.
solution: Offical solution to the problem, including LaTeX formatting… See the full description on the dataset page: https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset.formal-math-autoformalization
Formal Math Autoformalization Dataset
A growing, CC0 public-domain corpus of ⟨natural-language statement ↔ Lean 4 statement + proof⟩ pairs, contributed through the Agentic Commons network.
Why this is scarce data. Mathlib already contains millions of proven Lean theorems — but as bare Lean, with no paired natural language:
theorem add_comm (a b : ℕ) : a + b = b + a := ... -- no "addition on naturals is commutative" attached
The scarce, valuable artifact is the pairing of the… See the full description on the dataset page: https://huggingface.co/datasets/AgenticCommons/formal-math-autoformalization.Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1
Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1
Dataset Description:
Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1 is an RL dataset for training and evaluating a tool-using agent's ability to resist Indirect Prompt Injection (IPI) attacks hidden inside tool-returned environment data. In each record, the agent receives a benign user request that requires calling a read tool whose output contains an adversarial instruction disguised as legitimate domain content… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1.Nemotron-RL-Agentic-Function-Calling-Pivot-v1
Dataset Description:
This is a RL dataset for general function-calling by utilizing existing expert tool-use trajectories. We pose each assistant step of the trajectory as a separate behavior cloning problem where the policy model is incentivized to match the tool call choices of the expert model.
This dataset is released as part of NVIDIA NeMo Gym, a framework for building reinforcement learning environments to train large language models. NeMo Gym contains a growing collection… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1.Nemotron-RL-Agentic-Terminal-Pivot-v1
Dataset Description
The Nemotron-RL-Agentic-Terminal-Pivot-v1 dataset provides training samples for reinforcement learning of command-line ("terminal use") LLM agents with the terminus_judge environment in NeMo Gym.
Each record is a single agent decision point extracted from a successful agent trajectory on a terminal task:
responses_create_params.input — the prompt: the task instruction plus the terminal interaction history (prior agent actions and terminal outputs) up to the… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Terminal-Pivot-v1.Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1
Dataset Description:
We created an RL dataset for conversational tool-use by utilizing existing expert tool-use trajectories. We pose each assistant step of the trajectory as a separate behavior cloning problem where the policy model is incentivized to match the tool call choices of the expert model. Each trajectory includes the use of tools for authentication, data lookup, servicing (i.e. booking reservations, changing them, getting discounts, etc), and more across 838… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1.Nemotron-RL-Agentic-SWE-Pivot-v1
Dataset Description:
The SWE-RL dataset provides GitHub issues for training and validating real-world software engineering agents using the OpenHands environment in NeMo Gym. The dataset is a refactored version of the SWE-Gym and R2E-Gym datasets to support the NeMo Gym input format.
This dataset is released as part of NVIDIA NeMo Gym, a framework for building reinforcement learning environments to train large language models. NeMo Gym contains a growing collection of training… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-SWE-Pivot-v1.Agentic-SLS-ASTM
Agentic-SLS-ASTM
ASTM mechanical-test specimens (D638 tensile, D790 flex) printed on the Inova Mk1 SLS printer and pulled on an MTS / TestWorks Instron. Each row is a single specimen with full geometry, scalar results, stress–strain + raw DAQ curves, and — for SLS rows — FK references and an embedded snapshot of the upstream print profile from ppak10/Agentic-SLS-Database.
Rows are self-contained for ML use: the full PrintProfile JSON is inlined, so features (material/energy… See the full description on the dataset page: https://huggingface.co/datasets/ppak10/Agentic-SLS-ASTM.agentic-prompt-injection-boundary-pairs
Agentic Prompt-Injection Boundary Pairs
Most prompt-injection datasets make the attack easy to recognize. The malicious row contains obvious override language, while the benign row discusses something unrelated. A classifier can look capable without learning the boundary that matters in production.
This dataset takes a stricter approach. Each attack is paired with a legitimate request from the same workflow. The two rows share the asset, role, tool and topic. What changes is… See the full description on the dataset page: https://huggingface.co/datasets/3nesdeniz/agentic-prompt-injection-boundary-pairs.Agentic-Chain-of-Thought-Coding-SFT-Dataset
🤖 Agentic Coding CoT Dataset
A high-quality supervised fine-tuning (SFT) dataset for training agentic coding assistants with Chain-of-Thought reasoning capabilities.
📋 Dataset Description
This dataset was created by processing and distilling ~20GB of GitHub crawl data using Minimax-M2 to generate structured, reasoning-rich coding examples. Each sample demonstrates systematic problem-solving with explicit tool usage patterns.
🏗️ Assistant Data Structure… See the full description on the dataset page: https://huggingface.co/datasets/AlicanKiraz0/Agentic-Chain-of-Thought-Coding-SFT-Dataset.agentmujo-agentic-terminal
agentmujo-agentic-terminal (v0.1.0 — 541 trag)
Ručno dizajniran kanonski skup agentskih terminal-tragova na bosanskom
(ijekavica): dijagnoza → akcija → opservacija, dio
AgentMujo Training Frameworka.
Verzija: 0.1.0 · Tragova: 541 (split 450/42/49) · Jezik: bs-ijekavica
Format: JSONL kao function-calling skup; thinking uzorci nose
<think> reasoning blokove, svi nose tool_calls[] ili tekstualne odgovore.
Obrasci: jedan-korak statusi, višestepene dijagnoze (servis→mreža→store)… See the full description on the dataset page: https://huggingface.co/datasets/shaban2024/agentmujo-agentic-terminal.agentic_vbench_rollouts
agentic-vbench calibration rollout archive
Complete, immutable copies of the calibration trajectories referenced from the
agentic-vbench task PRs. Unlike the copies committed in each task's
calibration/rollouts/, nothing here is elided: every sampled-frame payload the agent
saw is present.
Redaction policy (mechanical, applied identically to every file):
Absolute host filesystem paths from the runner machine are replaced with /workspace.
Capture-hardware and source-collection… See the full description on the dataset page: https://huggingface.co/datasets/yalesunxiatao/agentic_vbench_rollouts.agentic-tool-call-dataset-12k
Agentic Tool Calling Dataset 12K
A curated 12K-sample tool-calling SFT dataset in a TRL-ready chat format. Each sample contains multi-turn agent trajectories with explicit reasoning, structured tool_calls, and tool responses.
Dataset Summary
Property
Value
Total Samples
12,000
Short split
10,000 (agent_short_10k.jsonl)
Long split
2,000 (agent_long_2k.jsonl)
Language
English
Format
OpenAI-style messages with tool_calls
License
Apache 2.0… See the full description on the dataset page: https://huggingface.co/datasets/pyromind/agentic-tool-call-dataset-12k.agentic-safety-gguf
agentic-safety-gguf: Training & Evaluation Datasets
Model: guerilla7/agentic-safety-ggufPaper: (https://arxiv.org/abs/2601.00848)Total: 80,992 examples (80,851 after deduplication)
Overview
Complete training and evaluation datasets for agentic-safety-gguf, a specialized Llama 3.1 8B model for agentic AI security analysis. Supports iterative continuation training methodology (V2→V3→V4) for full reproducibility.
Dataset Files
File
Examples
Size
Purpose… See the full description on the dataset page: https://huggingface.co/datasets/guerilla7/agentic-safety-gguf.agentic-llm-pretraining-1.7b
Agentic LLM Pretraining Dataset
A pretraining corpus for small language models (1-3B parameters) optimized for agentic tasks. The corpus emphasizes learning to comprehend language, reason, follow instructions, and use tools over memorizing factual knowledge — the assumption is that domain knowledge will be provided at runtime via RAG. The idea is that this could enable much smaller pretraining corpora by omitting the large volumes of text typically needed to memorize facts.… See the full description on the dataset page: https://huggingface.co/datasets/visionscaper/agentic-llm-pretraining-1.7b.agentic-sft-new
Agentic SFT Dataset
A comprehensive dataset for Agentic Supervised Fine-Tuning (SFT), curated and merged from multiple high-quality open-source datasets. It covers a wide range of agent capabilities including tool calling, code editing, terminal interaction, multi-hop reasoning, and web browsing.
Total samples: 711,852
Data Sources
This dataset is compiled from the following open-source datasets:
1. MiroVerse-v0.1 (147,985 samples)
Source:… See the full description on the dataset page: https://huggingface.co/datasets/WaltonFuture/agentic-sft-new.agentic_red_team
Agentic Red Team Tool-Calling Dataset
A multi-turn, tool-calling cybersecurity dataset where each example is a complete agentic trajectory — a realistic sequence of tool calls, tool responses, and reasoning steps that an AI agent would execute during an authorized red team engagement.
Overview
This dataset contains 5,000 agentic tool-calling examples across 20 offensive security sectors. Unlike traditional Q&A datasets, each row is a complete multi-turn trajectory… See the full description on the dataset page: https://huggingface.co/datasets/sh111111111111111/agentic_red_team.agentic-coding-tracesThis dataset is aggregated from:
"AletheiaResearch/GLM-5.2-Agent",
"armand0e/claude-fable-5-claude-code",
"armand0e/minimax-m3-claude-code-traces",
"AletheiaResearch/Kimi-K3-Codex"
and has been converted into a single unified format:
{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "<think>...</think>..."}, ...]}
The original raw files are kept in the .raw_sources folder for reference if you'd like to take a closer look.
Main file: USE_THIS_train.jsonl
ty
AgenticRAGTracer
AgenticRAGTracer: A Hop-Aware Benchmark for Diagnosing Multi-Step Retrieval Reasoning in Agentic RAG
Paper | Code
🎉 Our work has been accepted to ACL 2026 Findings!
AgenticRAGTracer is a benchmark designed to diagnose and evaluate multi-step retrieval reasoning in Agentic RAG systems. Unlike traditional benchmarks that provide only final questions and answers, AgenticRAGTracer includes intermediate hop-level questions that connect atomic questions to the final query. This allows… See the full description on the dataset page: https://huggingface.co/datasets/YqjMartin/AgenticRAGTracer.fable5-agentic-coding-sft
FABLE.5 Agentic Coding SFT (curated)
~159,972 supervised fine-tuning examples for agentic coding — multi-turn conversations where the
assistant drives a tool-call loop (shell, file edits, tests) and commits to complete solutions. Used to train
VibeThinker-Fable-Nano-Agentic-3B.
Provenance & license
Curated/distilled from the Complete-FABLE.5-traces-2M trace set:
Original source: Glint-Research/Complete-FABLE.5-traces-2M (currently gated).
Pulled from:… See the full description on the dataset page: https://huggingface.co/datasets/Nexlab/fable5-agentic-coding-sft.slimder-qwen38-agentic-ream-runs-20260901counsel-chat-miti-st-dpo
counsel-chat MITI-ST DPO
A preference-tuning (DPO) dataset for English psychology / counseling assistants,
derived from nbertagnolli/counsel-chat.
Each row is a (prompt, chosen, rejected) pair where both chosen and rejected
are real answers written by US licensed therapists to the same client question.
The LLM (Claude Opus 4.7) is used only as a scorer, never as a generator —
so the dataset does not contain any model-written counseling text, and DPO
training on it is not… See the full description on the dataset page: https://huggingface.co/datasets/AgenticCommons/counsel-chat-miti-st-dpo.agentic-rag-redteam-bench
WARNING: HARMFUL CONTENT - RESEARCH USE ONLY
This dataset contains adversarial prompts, jailbreak attacks, toxic outputs, and other explicitly harmful content generated for AI safety research. Samples include prompt injections, social engineering payloads, misinformation, hate speech, instructions for illegal activities, phishing templates, and other dangerous material. All content is synthetic and produced by automated red-teaming pipelines for the sole purpose of evaluating and improving… See the full description on the dataset page: https://huggingface.co/datasets/Fujitsu/agentic-rag-redteam-bench.agentic-asr
Agentic ASR
Public consolidated audio and ASR result dataset for the OSWorld and
WildClawBench benchmark families.
Layout
osworld/: synthetic raw/colloquial speech, human recordings, DNS-noise
pairs, task images, ASR results, and reports.
wildclawbench/: 60 formal colloquialized prompts, synthetic speech,
20 synthetic ASR condition tables, and ten-participant human recordings.
task0_template derivatives are excluded.
metadata/conditions.jsonl: model, variant… See the full description on the dataset page: https://huggingface.co/datasets/tterumiimurett1/agentic-asr.clawd-agentic-layer-whitepaper
Clawd Agentic Layer — Whitepaper v0.3
A design-proposal whitepaper for the Clawd Agentic Layer: a verifiable AI agent economy on Solana.
Abstract
Clawd is a sovereign Solana-native AI agent. This whitepaper specifies the Clawd Agentic Layer: onchain agent identity (Metaplex Agent Registry), Agent DNA lineage with law/constitution/proof hashes, a Six-Law Harness, a progressive trust ladder (Observer → Dry-Run → Delegated → Autonomous → Sovereign), a dual-wallet… See the full description on the dataset page: https://huggingface.co/datasets/ordlibrary/clawd-agentic-layer-whitepaper.repro-abc-bench-an-agentic-bio-capabilities-benchmark-for-biosecurity-traces
Agent traces
Agent sessions published from a Trackio Logbook.
hendar-agentic-ai-dataset
Hendar Agentic AI Evaluation & Security Benchmark
A compact, expert-authored benchmark for evaluating trustworthy agentic AI systems across capability, tool use, retrieval, security, policy enforcement, multi-agent coordination and regression safety. The current release contains 104 synthetic cases: 13 cases in each of 8 domains.
This dataset is a public companion to the Agentic AI Academy by Hendar Mawan, PhD. It is designed for evaluation, CI regression testing, red-team… See the full description on the dataset page: https://huggingface.co/datasets/h0000w/hendar-agentic-ai-dataset.hermes-agentic-tool-sft
Hermes Agentic Tool SFT
A 1,980-row synthetic SFT corpus for teaching open models to use
Hermes Agent tools and stop.
Measured problem (Muse Glimmer 30B, llama-server, DFlash off): a 20-task
simulated loop gate scored 7/20 with 7 HIT_CAP (12–14 tool calls,
empty answer). The same weights inside Hermes often do stop (native
4/5, mean 2.8 tools). This set trains the stop / don’t-retry behavior,
not generic “more tools.”
Companion bench: vcruz305/hermes-agentic-bench.… See the full description on the dataset page: https://huggingface.co/datasets/vcruz305/hermes-agentic-tool-sft.glm-5.2-nvfp4-agentic-traces
GLM-5.2 NVFP4 agentic software traces
This snapshot contains 1,989 completed Verifiers invocation records generated
with RedHatAI/GLM-5.2-NVFP4-FP8.
manifest.jsonl is a compact index for filtering and inspection.
data/<arm>.jsonl contains the exact full graph records emitted by Verifiers.
configs/<arm>.toml contains the resolved configuration for each arm.
The snapshot retains successes, failures, truncations, and scoring metadata.
Use solved, reward, has_error, failure_labels… See the full description on the dataset page: https://huggingface.co/datasets/synquid/glm-5.2-nvfp4-agentic-traces.lmcache-agentic-traces_Otel
Agentic LLM Traces – OTel Format
Overview
Real-world agentic LLM sessions converted to OpenTelemetry (OTel) trace format, derived from sammshen/lmcache-agentic-traces. Each session is a multi-turn agent interaction involving tool calls (bash commands, file edits, web search, etc.), spanning 5–50 turns and totalling 24,880 spans.
Traces come from three agentic benchmarks: SWE-bench, GAIA, and WildClaw. They are formatted as OTel spans following gen_ai.* semantic… See the full description on the dataset page: https://huggingface.co/datasets/ibm-research/lmcache-agentic-traces_Otel.