datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
kimi-k3-distillation
kimi-k3-distillation
Single-teacher slice of
r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation,
filtered to teacher_model == "kimi-code/k3" only. The Qwen3.8-Max-Preview and
GLM-5.2 traces are removed.
4,347 rows — 3,918 train / 212 validation / 217 test.
from datasets import load_dataset
ds = load_dataset("beyoru/kimi-k3-distillation") # sft: messages + tools
ds = load_dataset("beyoru/kimi-k3-distillation", "canonical") # + full audit columns… See the full description on the dataset page: https://huggingface.co/datasets/beyoru/kimi-k3-distillation.qwen3.8-max-glm5.2-kimi-k3-distillation
Multi-Teacher Distillation Dataset (57,937 traces)
A quality-filtered, deduplicated, multi-teacher SFT corpus combining traces from three frontier models across math, code, reasoning, instruction-following, tool-use, science, long-context, multilingual, and creative dialogue domains.
Teachers
Teacher
Provider
Traces
Qwen3.8-Max-Preview
Alibaba Cloud Model Studio
48,283
GLM-5.2
Z.AI Coding Plan
5,307
Kimi Code K3
Moonshot AI (Kimi)
4,347… See the full description on the dataset page: https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation.arc-agi3-kimi-k2.7-su15
ARC-AGI-3 su15 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game su15, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-su15.kimi-cyber-reasoning
Kimi Cyber Reasoning
997 chain-of-thought records covering 13 cybersecurity disciplines and 4 systems engineering domains, distilled from the Kimi K3 reasoning model via API. Every record provides an explicit step-by-step <think> reasoning trace followed by a technical resolution, unified code diff fix, or structured tool invocation.
The dataset was curated as an anchor set for training, healing, and specializing compact reasoning models on systems security and tool calling… See the full description on the dataset page: https://huggingface.co/datasets/echel0nn1881/kimi-cyber-reasoning.KIMI-K2.5-1000000x
KIMI-K2.5-1000000x
1,000,000 reasoning traces distilled from KIMI-K2.5 on high reasoning, (Each subset has different questions)
Distribution:
Coding: 50% (Includes: Webdev, Python, C++, Java, JS, C, Ruby, Lua, Rust, and C#)
Science: 20% (Physics, Chemistry, Biology) - 100k more completions in the PHD-Science subset
Math: 15% (Algebra, Calculus, Probability) - 200k more completions in kimiMath200k.jsonl
Computer Science: 5%
Logical Questions: 5%
Creative Writing: 5%… See the full description on the dataset page: https://huggingface.co/datasets/ianncity/KIMI-K2.5-1000000x.kimi-k3-coding-and-debugging-traces
Kimi K3 Coding, Tool Use & Instruction Following Traces
582 TRAJECTORIES · 3,956 TRAINING ROWS · 3 MB PARQUET · 72 MB JSONL
Generated by moonshiner — an open harness for
distilling verified instruction-following, tool-use, and agentic coding traces.
Behavior-preserving instruction-following, tool-use, and agent trajectories
from Kimi K3 (moonshotai/kimi-k3). The category and row-share tables
below describe the actual mix seen during training rather than assuming a… See the full description on the dataset page: https://huggingface.co/datasets/greghavens/kimi-k3-coding-and-debugging-traces.arc-agi3-kimi-k2.7-ls20
ARC-AGI-3 ls20 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game ls20, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-ls20.kimitobokunosaigonosenjouaruiwasekaigahajimaruseisenseasonii
Bangumi Image Base of Kimi To Boku No Saigo No Senjou, Aruiwa Sekai Ga Hajimaru Seisen Season Ii
This is the image base of bangumi Kimi to Boku no Saigo no Senjou, Aruiwa Sekai ga Hajimaru Seisen Season II, we detected 100 characters, 7344 images in total. The full dataset is here.
Please note that these image bases are not guaranteed to be 100% cleaned, they may be noisy actual. If you intend to manually train models using this dataset, we recommend performing necessary… See the full description on the dataset page: https://huggingface.co/datasets/BangumiBase/kimitobokunosaigonosenjouaruiwasekaigahajimaruseisenseasonii.AIME_2000_2026_Kimi_K3
AIME 2000–2026 — Kimi K3 reasoning traces
🔄 Changelog
2026-08-08 — full re-generation. All reasoning traces were regenerated from scratch and re-verified against the official answer key.
New schema — added gen_attempts_low, gen_attempts_high; renamed gen_parsed_answer → gen_answer_int and answer_note → problem_note; removed gen_effort, gen_pass1.
New generation — only use the bare problem (v1 appended an "ANSWER:" format instruction), so traces are cleaner.… See the full description on the dataset page: https://huggingface.co/datasets/bevangelista/AIME_2000_2026_Kimi_K3.arc-agi3-kimi-k2.7-g50t
ARC-AGI-3 g50t — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game g50t, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-g50t.qwen3.8-max-glm5.2-kimi-k3-distillation
Multi-Teacher Distillation Dataset (57,937 traces)
A quality-filtered, deduplicated, multi-teacher SFT corpus combining traces from three frontier models across math, code, reasoning, instruction-following, tool-use, science, long-context, multilingual, and creative dialogue domains.
Teachers
Teacher
Provider
Traces
Qwen3.8-Max-Preview
Alibaba Cloud Model Studio
48,283
GLM-5.2
Z.AI Coding Plan
5,307
Kimi Code K3
Moonshot AI (Kimi)
4,347… See the full description on the dataset page: https://huggingface.co/datasets/o0Biggz0o/qwen3.8-max-glm5.2-kimi-k3-distillation.arc-agi3-kimi-k2.7-tr87
ARC-AGI-3 tr87 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game tr87, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-tr87.arc-agi3-kimi-k2.7-ar25
ARC-AGI-3 ar25 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game ar25, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-ar25.discoverphysics-kimi-k2.7-ara
DiscoverPhysics × Kimi K2.7 (kimi-code CLI, thinking=on) — 11-world ARA knowledge artifacts
Agent-Native Research Artifacts (ARA) produced by a Kimi K2.7 (kimi-code CLI, thinking=on) coding-agent session solving all
11 worlds of the DiscoverPhysics scientific-discovery
benchmark (seed 0, noise_frac 0.075, ≤16 experiment rounds), driven through the same
harness-agnostic bridge and ARA scaffold as the sibling fable run. Official verdicts:
1/11 PASS — criteria and per-world numbers… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/discoverphysics-kimi-k2.7-ara.arc-agi3-kimi-k2.7-ft09
ARC-AGI-3 ft09 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game ft09, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-ft09.kiminokotogadaidaidaidaidaisukina100ninnokanojo
Bangumi Image Base of Kimi No Koto Ga Daidaidaidaidaisuki Na 100-nin No Kanojo
This is the image base of bangumi Kimi no Koto ga Daidaidaidaidaisuki na 100-nin no Kanojo, we detected 55 characters, 4768 images in total. The full dataset is here.
Please note that these image bases are not guaranteed to be 100% cleaned, they may be noisy actual. If you intend to manually train models using this dataset, we recommend performing necessary preprocessing on the downloaded dataset to… See the full description on the dataset page: https://huggingface.co/datasets/BangumiBase/kiminokotogadaidaidaidaidaisukina100ninnokanojo.kimi-k2.6-claude-code-tracesThis dataset was generated using teich by TeichAI
Prepare these datasets for supervised fine-tuning in just a few lines of code — see the Conversion section below.
Kimi K2.6 Claude Code Traces
This directory contains raw agent trace files generated by teich.
All assistant responses were generated by moonshotai/kimi-k2.6.
JSONL files: 36
Format
Each file is newline-delimited JSON representing a single captured agent session.
The trace schema is designed for… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/kimi-k2.6-claude-code-traces.kimi-k2.6-reap-observations-v1
Kimi-K2.6 REAP Observation Data (v1)
Per-layer expert routing + activation statistics captured from moonshotai/Kimi-K2.6
under the REAP layerwise observer (PR #17, CerebrasResearch/reap).
What this is
This dataset contains the observer output of a full REAP calibration pass on
Kimi-K2.6. It is not a pruned model. Each record describes per-token routing
decisions, expert activation norms, and the REAP saliency ingredients for every
MoE layer of the base model.
Downstream… See the full description on the dataset page: https://huggingface.co/datasets/mateowilliam/kimi-k2.6-reap-observations-v1.kimi-k3-distribution-fidelity-1024x2048-v1
Kimi K3 distribution-fidelity reference
Status
The artifact is qualified for paired, teacher-forced comparison of Kimi K3
compressed checkpoints against the official MXFP4 checkpoint. It contains
1,024 distinct 2,048-token source contexts and 2,096,128 scored next-token
positions.
The primary metric is:
KL(official MXFP4 distribution || candidate distribution)
The reference stores BF16 hidden states after final RMSNorm and before the
language-model head.… See the full description on the dataset page: https://huggingface.co/datasets/festr2/kimi-k3-distribution-fidelity-1024x2048-v1.kimiwameidosama
Bangumi Image Base of Kimi Wa Meido-sama.
This is the image base of bangumi Kimi wa Meido-sama., we detected 48 characters, 3792 images in total. The full dataset is here.
Please note that these image bases are not guaranteed to be 100% cleaned, they may be noisy actual. If you intend to manually train models using this dataset, we recommend performing necessary preprocessing on the downloaded dataset to eliminate potential noisy samples (approximately 1% probability).
Here is… See the full description on the dataset page: https://huggingface.co/datasets/BangumiBase/kimiwameidosama.kimi-k2.6-reap-observations-v1
Kimi-K2.6 REAP Observation Data (v1)
Per-layer expert routing + activation statistics captured from moonshotai/Kimi-K2.6
under the REAP layerwise observer (PR #17, CerebrasResearch/reap).
What this is
This dataset contains the observer output of a full REAP calibration pass on
Kimi-K2.6. It is not a pruned model. Each record describes per-token routing
decisions, expert activation norms, and the REAP saliency ingredients for every
MoE layer of the base model.
Downstream… See the full description on the dataset page: https://huggingface.co/datasets/0xSero/kimi-k2.6-reap-observations-v1.synthvision-validated-qwen-by-kimi
synthvision-validated-qwen-by-kimi
Qwen 3.5 annotations validated by Kimi K2.5 (93.1% pass rate)
Records: 55,359
About
Cross-validated subset from the SynthVision pipeline. Kimi K2.5 reviewed all 59,476 Qwen 3.5 annotations and confirmed 55,359 as consistent with the source images (93.1% pass rate).
Validation criteria: consistent == true AND confidence >= 0.7. Records that failed validation were removed — primarily cases where the annotator hallucinated findings not… See the full description on the dataset page: https://huggingface.co/datasets/OpenMed/synthvision-validated-qwen-by-kimi.qwen3.8-max-glm5.2-kimi-k3-distill
Multi-Teacher Distillation Dataset (57,937 traces)
A quality-filtered, deduplicated, multi-teacher SFT corpus combining traces from three frontier models across math, code, reasoning, instruction-following, tool-use, science, long-context, multilingual, and creative dialogue domains.
Teachers
Teacher
Provider
Traces
Qwen3.8-Max-Preview
Alibaba Cloud Model Studio
48,283
GLM-5.2
Z.AI Coding Plan
5,307
Kimi Code K3
Moonshot AI (Kimi)
4,347… See the full description on the dataset page: https://huggingface.co/datasets/ansulev/qwen3.8-max-glm5.2-kimi-k3-distill.qwen3.8-max-glm5.2-kimi-k3-distillation
Multi-Teacher Distillation Dataset (57,937 traces)
A quality-filtered, deduplicated, multi-teacher SFT corpus combining traces from three frontier models across math, code, reasoning, instruction-following, tool-use, science, long-context, multilingual, and creative dialogue domains.
Teachers
Teacher
Provider
Traces
Qwen3.8-Max-Preview
Alibaba Cloud Model Studio
48,283
GLM-5.2
Z.AI Coding Plan
5,307
Kimi Code K3
Moonshot AI (Kimi)
4,347… See the full description on the dataset page: https://huggingface.co/datasets/inferenceport-ai/qwen3.8-max-glm5.2-kimi-k3-distillation.Kimi-K2.5-Reasoning-1M-Cleaned
🪐 Kimi-K2.5-Reasoning-1M-Cleaned
Kimi-K2.5-Reasoning-1M-Cleaned is a cleaned derivative of ianncity/KIMI-K2.5-1000000x. It preserves the original four-config layout from the source dataset and rewrites each record into a unified reasoning-SFT schema with id, conversations, input, output, domain, and meta.
Summary
Source dataset: ianncity/KIMI-K2.5-1000000x
Source author: ianncity
Teacher model recorded in meta.teacher_model: KIMI-K2.5
Token lengths computed with… See the full description on the dataset page: https://huggingface.co/datasets/Jackrong/Kimi-K2.5-Reasoning-1M-Cleaned.arc-agi3-kimi-k2.7-s5i5
ARC-AGI-3 s5i5 — Agent Trajectories (kimi-k2.7)
Gameplay trajectories from the harness×model pair kimi-k2.7 playing the
ARC-AGI-3 game s5i5, part of the
ARA-as-world-model generalization experiment. The agent builds a structured world model
(an Agent-Native Research Artifact) live during play and consults it to crack levels it
cannot solve from cold exploration.
One dataset repo per harness×model×game: sibling repos
arc-agi3-<harness>-<model>-<game> hold the same game played by… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/arc-agi3-kimi-k2.7-s5i5.kimi_v31_300kfm-open-problems-kimi-k3-trajectories
FrontierMath Open Problems trajectories — kimi-k3
Fixed-budget agent trajectories on FrontierMath: Open Problems
(Epoch AI's collection of 50 genuinely unsolved research mathematics problems).
Nothing is graded. Epoch's verifiers are not public; the harness verifier is a
checkpoint stub that always writes reward 0 so the continue-until-timeout harness
keeps re-prompting the agent until the fixed wall-clock budget (105 min/task,
override_timeout_sec: 6300) elapses.… See the full description on the dataset page: https://huggingface.co/datasets/AgentNativeResearchLab/fm-open-problems-kimi-k3-trajectories.DeepSWE-Agent-Kimi-K2-Trajectories-2.8Kqwen3.8-max-glm5.2-kimi-k3-distillation
Multi-Teacher Distillation Dataset (57,937 traces)
A quality-filtered, deduplicated, multi-teacher SFT corpus combining traces from three frontier models across math, code, reasoning, instruction-following, tool-use, science, long-context, multilingual, and creative dialogue domains.
Teachers
Teacher
Provider
Traces
Qwen3.8-Max-Preview
Alibaba Cloud Model Studio
48,283
GLM-5.2
Z.AI Coding Plan
5,307
Kimi Code K3
Moonshot AI (Kimi)
4,347… See the full description on the dataset page: https://huggingface.co/datasets/bhadra123/qwen3.8-max-glm5.2-kimi-k3-distillation.