datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.SWE-Next
SWE-Next: Scalable Real-World Software Engineering Tasks for Agents
SWE-Next Dataset
SWE-Next is an execution-grounded dataset of 2,308 self-verifying software engineering tasks mined from real merged GitHub pull requests. Starting from 3,971 seeded Python repositories and 102,582 executed candidate base/merged commit pairs, SWE-Next retains only instances where the merged commit produces a strict test improvement without regressions. The final release… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/SWE-Next.qwen3.8-flash-next-expert-traces
Qwen3.8-Flash-Next expert routing traces
Token-level routing traces of a deployed MoE model: for every token and every one of the
48 MoE layers, which experts the router chose, the top-32 router logits behind that choice,
and the exact hidden state the router read — plus, in v3, the state at many layers per token,
the post-final-norm state the LM head consumes, and the LM head's top-8 next-token candidates.
The corpus exists to answer one question: how well can the next tokens'… See the full description on the dataset page: https://huggingface.co/datasets/aswinkumar99/qwen3.8-flash-next-expert-traces.agent-sft
Nex Agent-SFT Dataset
Paper | Code | Project Page
Dataset Description
This dataset is specifically designed for training the agentic capabilities of Large Language Models (LLMs). The dataset covers multiple agent scenarios and aims to enhance model performance in autonomous decision-making, tool usage, code generation, and interactive task handling. We reselected some of the training queries from the NEX-N1 training dataset and regenerated the responses based… See the full description on the dataset page: https://huggingface.co/datasets/nex-agi/agent-sft.swedish-legal-decisions-raw-v1
Swedish Court Decisions — Svenska Domstolsavgöranden
55,096 court decisions spanning 45 years of Swedish case law, purpose-built for LLM training.
The most comprehensive open dataset of Swedish appellate court decisions available for AI development. Sourced directly from the official Swedish Courts case law database via their public REST API and preprocessed into three ready-to-use training configurations.
Why This Dataset
Scale and depth: 55,096 decisions covering… See the full description on the dataset page: https://huggingface.co/datasets/nexoneAB/swedish-legal-decisions-raw-v1.NextCoderDataset
NextCoderDataset
GitHub | Paper
NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits (ICML'2025)
Data Overview
NextCoderdataset is the instruction-variant of synthetic dataset, used for training models on code-editing scenarios and compromised of around 381k (127k*3) samples across 8 different programming languages: Python,
Java, C++, C, Rust, Javascript, Go and Kotlin.
This is used to finetune the NextCoder family models using the novel… See the full description on the dataset page: https://huggingface.co/datasets/microsoft/NextCoderDataset.supra-nexus-o1-training
Supra Nexus O1 Training Datasets
Overview
Comprehensive training datasets for Supra Nexus O1 models, including:
Identity training
Chain-of-thought reasoning
Self-improvement examples (O1.5)
Instruction following
Datasets Included
1. Identity Dataset (supra_identity.jsonl)
Model identity and alignment
Organization information
Capability descriptions
2. Instruction Dataset (supra_instruct_*.jsonl)
Direct instruction… See the full description on the dataset page: https://huggingface.co/datasets/Supra-Nexus/supra-nexus-o1-training.SWE-Next-SFT-Trajectories
SWE-Next: Scalable Real-World Software Engineering Tasks for Agents
SWE-Next SFT Trajectories
SWE-Next SFT Trajectories is the supervised fine-tuning dataset released with SWE-Next: Scalable Real-World Software Engineering Tasks for Agents. It contains 3,693 ShareGPT-style multi-turn training examples collected from expert agent rollouts on 2,308 execution-grounded SWE tasks synthesized from real merged pull requests.
The dataset is designed for training… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/SWE-Next-SFT-Trajectories.task1729_personachat_generate_next
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task1729_personachat_generate_next
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks}… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task1729_personachat_generate_next.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.glm5-next-tiny-cpu-repro-v1This repository is an evidence bundle, not one root-format dataset at repository root.
first/ and repeat/ are separate complete sealed QFS root datasets; comparison/ holds
the comparison receipt and tokenwise result. panel/ is the sealed input panel. Other files
are provenance, logs and reproduction tools. Do not pass the bundle root as a QFS dataset.
GLM5-Next tiny native CPU fixture
This is a complete untrained random-initialized native Glm5NextForConditionalGeneration
wrapper… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm5-next-tiny-cpu-repro-v1.Qwen3-Coder-Next-Open-Code-SFT
Dataset Card — OpenCode Rejection Sampling
Overview
This dataset contains high-quality code reasoning data for training language models on competitive programming tasks. It is produced via rejection sampling with Qwen3-Coder-Next, which would generate multiple candidate solutions per problem, each candidate is executed against test cases in a sandboxed environment, and the results are used to build two complementary training datasets:
SFT dataset (49,374 examples)… See the full description on the dataset page: https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-Open-Code-SFT.Qwen3-Coder-Next-OpenCode-Preference
Dataset Card — OpenCode Rejection Sampling (Preference)
Overview
This dataset contains 10,920 preference pairs for preference-based training (DPO, KTO, SimPO, ORPO, etc.) on competitive programming tasks. Each pair consists of:
Chosen: a candidate solution that passes 100% of test cases
Rejected: a candidate solution that fails, with a fine-grained rejection type label
Pairs are produced via rejection sampling with Qwen3-Coder-Next: 8 candidate solutions are… See the full description on the dataset page: https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-OpenCode-Preference.Medical-Reasoning-SFT-Qwen3-Next-80B
Medical-Reasoning-SFT-Qwen3-Next-80B
A large-scale medical reasoning dataset generated using Qwen/Qwen3-Next-80B-A3B-Thinking, containing over 604,000 samples with detailed chain-of-thought reasoning for medical and healthcare questions.
Dataset Overview
Metric
Value
Model
Qwen/Qwen3-Next-80B-A3B-Thinking
Total Samples
604,249
Samples with Reasoning
604,249 (100%)
Estimated Tokens
~1.42 Billion
Content Tokens
~505 Million
Reasoning Tokens
~917 Million… See the full description on the dataset page: https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Qwen3-Next-80B.terminal-wm-sft-nextobs-v2
Terminal World-Model SFT (nextobs, corrected v2)
Corrected supervised fine-tuning data derived from
open-thoughts/OpenThoughts-Agent-SFT-100K
Terminus traces. Schema version: v2-observation-action.
Causal format
Every world-model transition is serialized as:
system: target-specific world-model instruction
user: task context (turn 1) + real current observation_t + executed action_t
assistant: target derived from the real observation_t+1
Rows remain multi-turn… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/terminal-wm-sft-nextobs-v2.nextbench
NextBench
The benchmark for modern Next.js code generation and completion.
NextBench measures how well a language model can complete real-world Next.js / React / TypeScript code. Every task is an autocomplete prompt — a partial file with the cursor at the end — graded against deterministic checks: must-contain patterns, forbidden patterns, regex matches, and output length.
443 tasks across 16 categories (v0.2)
12-model leaderboard spanning 1.3B–30B parameters; reproducible from… See the full description on the dataset page: https://huggingface.co/datasets/baablabs/nextbench.terminal-wm-sft-nextobs
Terminal World-Model SFT data — nextobs target
Supervised fine-tuning data for training a Terminal World Model: given the interaction history and the
current keystroke action, predict the next terminal observation. Built from
open-thoughts/OpenThoughts-Agent-SFT-100K
(terminus-2 keystroke traces in containerized Linux shells) via recipe/wm_rl/wm_sft/build_sft_dataset.py.
This variant's target
The assistant turn (the SFT target) is the raw next terminal observation… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/terminal-wm-sft-nextobs.NextSearch-1-Trajectories
NextSearch-1 Trajectories
The supervised training trajectories behind the
NextSearch-1 web research agents:
complete research episodes — reasoning, tool calls, live-web tool results,
and final answers — for every task in the companion
NextSearch-1-Tasks
SFT configs. Directly trainable: each row is a prompt (messages) plus a
target trajectory (target) with per-message reasoning and OpenAI-format
tool calls.
Technical report: nexttoken.co/research/nextsearch-1 ·
Harness and evals:… See the full description on the dataset page: https://huggingface.co/datasets/NextTokenAI/NextSearch-1-Trajectories.task963_librispeech_asr_next_word_prediction
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task963_librispeech_asr_next_word_prediction
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task963_librispeech_asr_next_word_prediction.tat-llm-instructions
TAT-LLM-Instructions
The TAT(Tabular and Textual)-LLM-Instructions dataset is a curated collection of financial data, structured to resemble instructions. It aggregates information from three publicly available tabular and textual QA datasets: FinQA, TAT-QA, and TAT-DQA. By employing specialized templates, TAT-LLM-Instructions transforms the original dataset into prompts that are optimized for compatibility with large language models (LLMs) and external executor, aiming to… See the full description on the dataset page: https://huggingface.co/datasets/next-tat/tat-llm-instructions.next.js-15.4-with-reasoning
Description
The Next.js Documentation Dataset based on next.js 15.4 version is a high-quality, code-centric dataset created from Next.js documentation for fine-tuning language models. It contains 1,172 question-answer pairs derived from 178 markdown documentation files, focusing on practical code examples and real-world development scenarios.
This dataset is designed for:
Question Answering: Natural language questions about Next.js development
Code Generation: Generating practical… See the full description on the dataset page: https://huggingface.co/datasets/Slava32/next.js-15.4-with-reasoning.NextCoderDataset-Conversational
NextCoderDataset-Conversational
GitHub | Paper
NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits (ICML'2025)
Data Overview
NextCoderDataset-Conversational is the multi-turn conversational-variant of synthetic dataset, used for training models on code-editing scenarios and compromised of around 57k samples across 8 different programming languages: Python,
Java, C++, C, Rust, Javascript, Go and Kotlin.
This is used to finetune the NextCoder… See the full description on the dataset page: https://huggingface.co/datasets/microsoft/NextCoderDataset-Conversational.omnimcp_nextjs_react_architect_teaser
🔬 INSPECT THE DEEPSEEK-R1 REASONING CHAIN LIVE:
Zero hallucinations. Null syntax errors. 100% AST compiler validated.🌐 Live Interactive Reasoning & Code Inspector: https://emgena.com/trainingslager🎁 Claim your Free Starter Kit (Code: STARTER100): https://emgena.com/trainingslager🏷️ Launch Discount: Get 20 € OFF any 500-incident production suite with code LAUNCH20!
📜 Enterprise Compliance: EU AI Act Articles 50 & 53 certified • 100% DSGVO / GDPR clean • Commercial EULA… See the full description on the dataset page: https://huggingface.co/datasets/emgena/omnimcp_nextjs_react_architect_teaser.socraticDataset1680
Socratic AI Pedagogy Preference Dataset (1,680 Quadruplets)
A curated multi-domain educational preference dataset for post-training open language models into pedagogical Socratic tutors for introductory Artificial Intelligence and Machine Learning courses.
📚 Dataset Overview
The dataset contains 1,680 paired preference quadruplets across five foundational AI subfields:
Classical Search & Planning (347 items): A* heuristic admissibility, graph search state-space… See the full description on the dataset page: https://huggingface.co/datasets/NextGenInstitute/socraticDataset1680.NextSearch-1-Tasks
NextSearch-1 Tasks
The task pools behind the NextSearch-1
web research agents: every row is a research question with its reference
answer and grading spec — the sft-tasks configs are the tasks behind the
supervised corpora, the rl-tasks configs the verified prompt+gold pools
used for reinforcement learning. Full trajectories for the SFT configs are in the companion
NextSearch-1-Trajectories.
Technical report: nexttoken.co/research/nextsearch-1 ·
Harness and evals:… See the full description on the dataset page: https://huggingface.co/datasets/NextTokenAI/NextSearch-1-Tasks.Qwen3-Coder-Next-Open-Code-SFT
Dataset Card — OpenCode Rejection Sampling
Overview
This dataset contains high-quality code reasoning data for training language models on competitive programming tasks. It is produced via rejection sampling with Qwen3-Coder-Next, which would generate multiple candidate solutions per problem, each candidate is executed against test cases in a sandboxed environment, and the results are used to build two complementary training datasets:
SFT dataset (49,374 examples)… See the full description on the dataset page: https://huggingface.co/datasets/derekib/Qwen3-Coder-Next-Open-Code-SFT.omnimcp_nextjs_server_actions_teaser
🔬 INSPECT THE DEEPSEEK-R1 REASONING CHAIN LIVE:
Zero hallucinations. Null syntax errors. 100% AST compiler validated.🌐 Live Interactive Reasoning & Code Inspector: https://emgena.com/trainingslager🎁 Claim your Free Starter Kit (Code: STARTER100): https://emgena.com/trainingslager🏷️ Launch Discount: Get 20 € OFF any 500-incident production suite with code LAUNCH20!
📜 Enterprise Compliance: EU AI Act Articles 50 & 53 certified • 100% DSGVO / GDPR clean • Commercial EULA… See the full description on the dataset page: https://huggingface.co/datasets/emgena/omnimcp_nextjs_server_actions_teaser.reasoning-sft-NextCoderDataset-100K
NextCoderDataset (converted)
Converted version of microsoft/NextCoderDataset, subsampled to 100,000 rows equally distributed across 8 programming languages for reasoning SFT training.
Format
Each row has three columns:
input - list of dicts with system and user messages (system prompt sets expert code editor role, user prompt contains the editing instruction and original code)
response - response string with <think> reasoning block followed by the edited code in markdown… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/reasoning-sft-NextCoderDataset-100K.nexus-sft-v1-sources
nexus-sft-v1-sources
AETHER family SFT dataset — group nexus_core.
Format: JSONL ChatML messages, task_type tagged, MinHash dedup applied (threshold 0.85).
Schema:
{
"messages": [{"role": "system|user|assistant", "content": "..."}],
"task_type": "function_calling|code|reasoning_cot|...",
"source_ds": "<HF dataset_id>",
"lang": "en|fr|...",
"system_source": "archon_default|overridden_from_source"
}
Generated by prepare_sft.py pipeline (2026-05-25).
llm-agent-harness-reliability-next-prime
LLM Next Prime Harness Dataset
This dataset contains raw observations from an experiment studying how the agent harness affects reliability when an LLM has access to a deterministic tool.
The task is deliberately simple and objectively verifiable:
What is the smallest prime number that is strictly greater than n?
The deterministic tool computes the correct answer with a local Python next_prime(n) function. The experiment asks whether failures come from the model, the provider… See the full description on the dataset page: https://huggingface.co/datasets/hoololi/llm-agent-harness-reliability-next-prime.