Datasets

← Home

116 datasets found

Filter by tag: sft (36) · reasoning (25) · distilled (24) · agentic (21) · tool-calling (20) · coding (19) · math (8) · open-harness (8) · nvidia (6) · code (6) · cot (6) · deepseek-v4 (5) · deepseek-r1 (5) · grpo (5) · instruction-tuning (5) · non-commercial (5) · pretraining (4) · synthetic (4) · web-text (4) · stem (4) · multilingual (4) · glm (3) · zhipu (3) · rl (3) · dpo (3) · function-calling (3) · tiger-lab (3) · huggingface (3) · swe (3) · multi-turn (3)

Sort: popular · downloads · stars · newest

• apigen-function-calling — This dataset is a merge of [argilla/Synth-APIGen-v0.1](https://huggingface.co/datasets/argilla/Synth-APIGen-v0.1) and [Salesforce/xlam-function-calling-60k](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k), making over 100K function calling examples following the APIGen recipe. (0 downloads, 0 stars, cc-by-4.0, )
• tool_calling — BitAgent tool_calling: synthetic function-calling conversations from the BitAgent project (BerryML) covering multi-tool selection, nested arguments, and rejection cases. No license tag upstream (default HF terms). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• xlam_hermes_validated — xlam-Hermes validated: the Salesforce xlam function-calling corpus re-formatted to the Hermes tool-use template and filtered for validity by Interstellar Ninja — pairs with the xlam/Hermes function-calling sets already mirrored. No license tag upstream (default HF terms). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• tool-use-multiturn-reasoning — --- dataset_info: features: - name: conversations list: - name: from dtype: string - name: value dtype: string - name: tools dtype: string - name: task dtype: string - name: category dtype: string - name: source dtype: string splits: - name: train num_bytes: 138012926 num_examples: 14579 download_size: 26612632 dataset_size: 138012926 configs: - config_name: default data_files: - split: train path: data/train-* license: apache-2.0 task_categories: - question-answering language: - en tags: - tool-use - multiturn - agentic - bfcl - reasoning size_categories: - 10K<n<100K --- (0 downloads, 0 stars, apache-2.0, )
• tool-calls-singleturn — Tool-calls single-turn: Hermes-format single-turn function-calling samples (system prompt + tool schemas + one assistant tool-call response) from the Interstellar Ninja tool-use family. Apache-2.0. (0 downloads, 0 stars, apache-2.0, )
• RLVR-IFeval — RLVR-IFeval (AllenAI): instruction-following tasks with programmatically verifiable constraints (from IFEval) for Tulu-3 RLVR training — prompts carry machine-checkable requirements so rewards can be computed without a reward model. ODC-By. (0 downloads, 0 stars, odc-by, )
• openchat_sharegpt4_dataset — This repository contains cleaned and filtered ShareGPT GPT-4 data used to train OpenChat. Details can be found in the [OpenChat repository](https://github.com/imoneoi/openchat). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• no_robots — - **Repository:** https://github.com/huggingface/alignment-handbook - **Paper:** - **Leaderboard:** https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard - **Point of Contact:** Lewis Tunstall (0 downloads, 0 stars, cc-by-nc-4.0, )
• OpenOrca — OpenOrca (1M GPT-4-augmented split): the augmented FLAN collection — ~1M GPT-4 completions over FLAN prompts (plus 3.2M GPT-3.5 in the other file), the open analogue of the Orca paper's system-messages approach. MIT. (0 downloads, 0 stars, mit, )
• ultrachat_200k — UltraChat 200k (HuggingFaceH4): the filtered UltraChat corpus used to train Zephyr-7B — ~200k multi-turn open-ended dialogues generated against GPT-3.5-Turbo across ~30 domains, decontaminated against eval benchmarks. MIT. (0 downloads, 0 stars, mit, )
• COIG-CQIA — COIG-CQIA (m-a-p): the Chinese Open Instruction Generalist quality-filtered split — ~48k high-quality Chinese instruction sources spanning academia, government, exams, and media, widely used for Chinese-language SFT. No license tag upstream (default HF terms). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• MATH-plus — This dataset contains the MetaMath, MATH-orca and some additional MATH-augmented dataset with GPT-4. This dataset is being used to train MAmmoTH2-plus version (https://tiger-ai-lab.github.io/MAmmoTH2/). (0 downloads, 0 stars, mit, )
• code_contests_instruct — The `deepmind/code_contests` dataset formatted as markdown-instruct for text generation training. (0 downloads, 0 stars, apache-2.0, )
• SWE-Bench-Verified-O1-native-tool-calling-reasoning-high-results — This dataset was generated using the CodeAct framework, which aims to improve code generation through enhanced action-based reasoning. Built on top of OpenHands, a framework designed for multi-turn interactive programming tasks, we tested O1 issue resolution capabilities on ```reasoning_effort = 'high'``` (0 downloads, 0 stars, cc-by-4.0, )
• hf-coding-tools-traces — This dataset rehydrates the benchmark run in [`davidkling/hf-coding-tools-dashboard`](https://huggingface.co/buckets/davidkling/hf-coding-tools-dashboard) into the JSONL session format consumed by the [Hugging Face Agent Trace Viewer](https://huggingface.co/changelog/agent-trace-viewer). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• transformers-coding-session-pi-traces — pi coding-agent session traces produced by [agentcap](https://github.com/huggingface/agentcap) runs. Each run contributes one folder under `data/<run_id>/`; inside, one file per session in `pi`'s native export format. (0 downloads, 0 stars, apache-2.0, )
• agentic-coding-traces — This 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. (0 downloads, 0 stars, apache-2.0, )
• real-pi-coding-agent-traces-sessions — An aggregated dataset of **real human–AI coding agent sessions**, collected from 21 independently published Hugging Face datasets and hand-filtered to exclude synthetic or AI-generated content. (0 downloads, 0 stars, other, )
• swe-rebench-v2-glm-5.1-pi-agent-successful-traces — We collected 7,777 successful multi-turn trajectories by running [GLM-5.1](https://huggingface.co/zai-org/GLM-5.1) with the [PI agent](https://pi.dev/) on [nebius/SWE-rebench-V2](https://huggingface.co/datasets/nebius/SWE-rebench-V2). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• swesmith-clean — swesmith-clean: a cleaned variant of the SWE-smith corpus from the R2E-Gym team — synthetically-injected SWE tasks with filtered execution results, for training and evaluating repository-level coding agents. No license tag upstream (default HF terms). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• Agentic-Multi-SWE-RL — Agentic Multi-SWE-RL: multi-language software-engineering RL trajectories from II-Vietnam — agent episodes across SWE-bench-style multilingual (Java/TypeScript/Go/Rust…) repo tasks with verifiable outcomes, for GRPO/RLVR training of SWE agents. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• SWE-Factory-Gym — SWE-Factory-Gym: software-engineering agent training environments generated by the SWE-Factory pipeline — repo-level task instances with executable test harnesses for SFT and verifiable-reward (RLVR) agent training. MIT. (0 downloads, 0 stars, mit, )
• SWE-Gym — SWE-Gym contains 2438 instances sourced from 11 Python repos, following SWE-Bench data collection procedure. (0 downloads, 0 stars, mit, )
• OpenHands-SFT-Trajectories — --- dataset_info: features: - name: messages list: - name: content dtype: string - name: role dtype: string splits: - name: train.success.oss num_bytes: 31805636 num_examples: 491 download_size: 10379409 dataset_size: 31805636 configs: - config_name: default data_files: - split: train.success.oss path: data/train.success.oss-* license: mit --- (0 downloads, 0 stars, mit, )
• SWE-smith-trajectories — SWE-smith trajectories (official, SWE-bench org): execution trajectories on the SWE-smith synthetic task corpus — model attempts with tool usage and patch outcomes across 32 parquet shards, the training substrate for SWE-agent-style models. MIT. (0 downloads, 0 stars, mit, )
• SWE-agent-trajectories — The dataset collection consisted of two stages: collecting issue-resolution instances, following a methodology similar to SWE-bench, and generating a large number of trajectories for solving the collected issues. The generated code patches in these trajectories were evaluated by the tests from the linked pull requests to determine which of them passed the tests. The detailed process of collecting issue-resolution instances is described in our blog post ["Scaling Data Collection for Training Software Engineering Agents"](https://nebius.com/blog/posts/scaling-data-collection-for-training-swe-agents). (0 downloads, 0 stars, cc-by-4.0, )
• SWE-smith-mini_swe_agent_plus-trajectories-66k — SWE-smith mini-swe-agent trajectories (Kwai-Klear): 66k agent trajectories collected with mini-swe-agent-plus on SWE-smith tasks — repository-level issue resolution episodes with tool calls, for SFT and RL of SWE agents. MIT. (0 downloads, 0 stars, mit, )
• codeforces-submissions — This dataset includes millions of real user (human) code submissions to the CodeForces website. (0 downloads, 0 stars, cc-by-4.0, )
• Big-Math-RL-Verified-Processed — This is a processed version of [SynthLabsAI/Big-Math-RL-Verified](https://huggingface.co/datasets/SynthLabsAI/Big-Math-RL-Verified) where we have applied the following filters: (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• OpenThoughts-114k-math — This is a filtered and metadata enriched version of [`open-thoughts/OpenThoughts-114k`](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• DAPO-Math-17k-Processed — This is a processed version of [BytedTsinghua-SIA/DAPO-Math-17k](https://huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k) where we have: (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• AM-Thinking-v1-Distilled — AM-Thinking-v1 and Qwen3-235B-A22B are two reasoning datasets distilled from state-of-the-art teacher models. Each dataset contains high-quality, automatically verified responses generated from a shared set of **1.89 million queries** spanning a wide range of reasoning domains. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), )
• Nemotron-Terminal-Corpus — Terminal-Corpus is a large-scale Supervised Fine-Tuning (SFT) dataset designed to scale the terminal interaction capabilities of Large Language Models (LLMs). Developed by NVIDIA, this dataset was built using the **Terminal-Task-Gen** pipeline, which combines dataset adaptation with synthetic task generation across diverse domains. (0 downloads, 0 stars, cc-by-4.0, )
• Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1 — 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 different domains such as EdTech, IT support, travel, etc. (0 downloads, 0 stars, cc-by-4.0, )
• DeepCoder-Preview-Dataset — - 7.5K **TACO Verified** problems. - 16K verified coding problems from **PrimeIntellect’s SYNTHETIC-1.** - 600 **LiveCodeBench (v5)** problems submitted between **May 1, 2023** and **July 31, 2024.** (0 downloads, 0 stars, mit, )
• DeepScaleR-Preview-Dataset — Our training dataset consists of approximately 40,000 unique mathematics problem-answer pairs compiled from: (0 downloads, 0 stars, mit, )
• HelpSteer2 — HelpSteer2 is an open-source Helpfulness Dataset (CC-BY-4.0) that supports aligning models to become more helpful, factually correct and coherent, while being adjustable in terms of the complexity and verbosity of its responses. This dataset has been created in partnership with [Scale AI](https://scale.com/). (0 downloads, 0 stars, cc-by-4.0, )
• AM-Qwen3-Distilled (code) — Qwen3-distilled coding/reasoning traces (~15 GB code split; math/IF splits also in repo). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), coding, reasoning, qwen3)
• AM-DeepSeek-R1-Distilled (0.5M slice) — 500k-row DeepSeek-R1-distilled reasoning slice (of 1.4M) — NON-COMMERCIAL (CC-BY-NC-4.0). (0 downloads, 0 stars, non-commercial research only (cc-by-nc-4.0), reasoning, cot, deepseek-r1)
• Verifiable Coding Problems (Python) — Python problems + executable test cases for verifiable-reward (RLVR/GRPO) coding training. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), coding, rlvr, grpo)
• OpenR1 Codeforces CoTs (solutions) — Codeforces problems with R1-style CoT solutions (competitive programming, CC-BY-4.0). (0 downloads, 0 stars, cc-by-4.0, coding, competitive-programming, reasoning)
• Mixture-of-Thoughts (open-r1) — Curated R1 reasoning traces across math/code/science (open-r1's MoT recipe). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), reasoning, cot, deepseek-r1)
• LongWriter-6k — 6k samples with 2k–32k-word outputs — the long-form writing recipe (Apache-2.0). (0 downloads, 0 stars, apache-2.0, long-form, writing, sft)
• LongAlign 10K — 10k long-context instructions (4k–192k tokens) from Zhipu's LongAlign recipe. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), long-context, sft, glm)
• AgentInstruct (Zhipu) — Zhipu's multi-environment agent trajectories (ALFWorld/WebShop/Mind2Web/…) — the classic agentic SFT set. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), agentic, computer-use, tool-calling)
• Hermes Reasoning Tool-Use — Hermes-format interleaved reasoning + tool-call trajectories (Apache-2.0) — pairs with Hermes FC v1. (0 downloads, 0 stars, apache-2.0, tool-calling, reasoning, cot)
• WebInstructSub — 2.3M mined STEM QA pairs (forums + StackExchange) from the MAmmoTH2 pipeline (Apache-2.0). (0 downloads, 0 stars, apache-2.0, stem, sft, web-mined)
• MathInstruct — 262k math reasoning traces (CoT + Program-of-Thought) from 13 sources (MIT). (0 downloads, 0 stars, mit, stem, math, sft)
• Orca-Math Word Problems 200K — 200k math word problems with GPT-4 answers (the Orca-Math recipe, MIT). (0 downloads, 0 stars, mit, stem, math, sft)
• Evol-Instruct-Code 80K v1 — 80k GPT-4-evolved coding instructions (the WizardCoder corpus) — NON-COMMERCIAL (CC-BY-NC-SA-4.0). (0 downloads, 0 stars, non-commercial research only (cc-by-nc-sa-4.0), coding, sft, evol-instruct)