Distillio — the open source training data collective

Distillio is a community-run library of LLM training datasets (RLHF, instruction tuning, math, code, conversation), a curated index of open-weight models for fine-tuning and local GGUF inference, and free fine-tuning notebooks. Browsing and downloading are open to everyone; posting, mirroring, and starring require a free account.

• Browse datasets
• Submit a dataset
• Model hub (fine-tunable safetensors + local GGUF quants)
• Free fine-tuning notebooks

Platform stats

• 116 datasets
• 23 downloads
• 0 mirrors
• 16 stars
• 7 researchers

Trending datasets

• Qwen3.8-Max + GLM-5.2 + Kimi K3 Multi-Teacher Distillation — 57,937 filtered multi-teacher SFT traces from Qwen3.8-Max, GLM-5.2, and Kimi K3 across 24 configs with per-row provenance. (0 downloads, 0 stars, non-commercial research only (mixed upstream: CC BY-NC + Alibaba Cloud Model Studio terms), sft, distilled, multi-teacher)
• kimi-k3-coding-and-debugging-traces — Agentic coding/tool-use trajectories from Kimi K3, generated via moonshiner. (4 downloads, 3 stars, unspecified (source: Kimi K3, custom Kimi K3 License — permits derivatives, no distillation restriction), tool-calling, agentic, coding)
• DeepSeek-V4-Pro-Distilled-200K — 188k long-form math & STEM reasoning traces distilled from DeepSeek-V4-Pro in Max mode. (0 downloads, 0 stars, cc-by-4.0 + cc-by-sa-4.0, reasoning, distilled, math)
• Llama-Nemotron-Post-Training-Dataset — NVIDIA's 30M+ example SFT+RL post-training dataset behind Llama-3-Nemotron. (0 downloads, 0 stars, cc-by-4.0, sft, reinforcement-learning, reasoning)
• xlam-function-calling-60k — 60k function-calling examples across 3,673 APIs via Salesforce's APIGen. (0 downloads, 0 stars, mixed signal (cited as cc-by-4.0, but source repo says research-only; gated), tool-calling, function-calling, agentic)
• OpenAssistant Conversations (oasst1) — Crowdsourced human-written multi-turn conversation trees from Open Assistant. (1 downloads, 0 stars, apache-2.0, sft, human-written, conversations)
• Deepseek-v4.1-CoT — Chain-of-thought data from DeepSeek-V4.1-Flash, released explicitly for distillation. (0 downloads, 0 stars, apache-2.0, reasoning, distilled, cot)
• DeepSeek v4 Pro Agent Traces — 4,006 raw agentic session traces from DeepSeek-V4-Pro via the open teich harness. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), agentic, tool-calling, distilled)
• GLM-5.2 Agent traces — 319 raw agentic session traces from GLM-5.2 via the open teich harness, with recovered tool schemas. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), agentic, tool-calling, distilled)
• GLM 5.2 Agent Traces (moonshiner) — Behavior-preserving GLM-5.2 instruction-following, tool-use, and agentic coding trajectories from moonshiner. (0 downloads, 0 stars, cc-by-4.0, tool-calling, agentic, coding)
• Kimi Cyber Reasoning — 997 Kimi K3 chain-of-thought records across 13 cybersecurity disciplines with tool-calling records. (0 downloads, 0 stars, wtfpl, reasoning, distilled, cybersecurity)
• Fable-5-Max Reasoning (Nanbeige filter) — 128 high-density fintech security-architecture traces distilled from Fable-5-Max (eBPF, Istio, SPIFFE/Vault). (0 downloads, 0 stars, apache-2.0 (teacher: Fable-5-Max — check Anthropic usage terms for your use), reasoning, distilled, security)
• DeepSeek-V4-Distill-8000x — Reasoning SFT distilled from DeepSeek-V4-Flash as teacher. (12 downloads, 5 stars, unspecified (source: DeepSeek-V4-Flash, MIT + explicit distillation permission), reasoning, distilled, deepseek-v4)
• SmolTalk2 — The 2025 SFT mixture behind SmolLM3: decontaminated multi-task instruction data spanning chat, reasoning, and tool use. (0 downloads, 0 stars, mixed (apache-2.0 + source licenses), sft, instruction-tuning, multi-task)
• SmolTalk — Curated Magpie-Ultra-based instruction mix used to train the SmolLM2 family. (0 downloads, 0 stars, mixed (apache-2.0 + source licenses), sft, instruction-tuning, magpie)
• Tulu 3 SFT Mixture — AI2's ~939k-example multilingual SFT corpus with full provenance labels. (1 downloads, 0 stars, odc-by, sft, instruction-tuning, multilingual)
• UltraFeedback (binarized) — Binarized chosen/rejected preference pairs from UltraFeedback, ready for DPO. (0 downloads, 0 stars, mit, preference, dpo, rlhf)
• Fable-5 Coding & Debugging Traces (Synthetic Corrections) — Synthetically corrected, re-judged Fable-5 coding trajectories that failed in the original set (tiny companion set). (0 downloads, 0 stars, cc-by-4.0 (teacher: Claude Fable 5 — check Anthropic usage terms for your use), coding, agentic, distilled)
• Nemotron Terminal Corpus (SWE adapter) — Terminal SWE sessions from NVIDIA's Terminal Corpus (swe adapter; code/math adapters also in repo). (0 downloads, 0 stars, cc-by-4.0, agentic, terminal, coding)
• Falcon RefinedWeb — ~600B-token filtered, deduplicated CommonCrawl extract used to pretrain Falcon. (0 downloads, 0 stars, apache-2.0, pretraining, web-text, filtered)
• GPT-5.6 Sol Coding & Debugging Traces — Verified GPT-5.6 Sol agentic coding, debugging, and defensive-security trajectories via Codex CLI. (0 downloads, 0 stars, cc-by-4.0 (teacher: GPT-5.6 Sol — check OpenAI terms for your use), coding, agentic, distilled)
• Qwen3.8-Max Distillation 50K — 49,772 teacher traces from Qwen3.8-Max-Preview for SFT and off-policy distillation (non-commercial). (0 downloads, 0 stars, non-commercial research only (CC BY-NC sources + Alibaba Cloud Model Studio terms), reasoning, distilled, math)
• Meta Muse Spark Distilled 5K — 5,000 synthetic Understand→Plan→Execute→Verify reasoning traces in 10 categories (style-mirror, not model output). (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), reasoning, synthetic, cot)
• CAD-1K Open v2 (cad-1000-hours) — 509 recorded CAD tasks / 1,018 hours with rubrics, screen recordings, and deliverables across 7 CAD suites. (0 downloads, 0 stars, unspecified (community upload, no license tag — default HF terms), agentic, computer-use, evaluation)
• Argilla Intel Orca DPO Pairs — DPO preference pairs distilled from the Intel Orca instruction set via Argilla's distilabel pipeline (Apache-2.0). (0 downloads, 0 stars, apache-2.0, preference, dpo, orca)
• NVIDIA Open-SWE Traces (scale-swe, Qwen3.6-27B) — SWE agent trajectories (mini-swe-agent, Qwen3.6-27B, scale-swe slice): 17 parquet shards of repo-level issue solving. (0 downloads, 0 stars, cc-by-4.0, agentic, coding, swe)
• RedPajama-Data-1T — Together AI's 1.2T-token open reproduction of the LLaMA pretraining mix. (0 downloads, 0 stars, apache-2.0, pretraining, web-text, open-reproduction)
• Dolma 3 Pool — AI2's 9T-token pretraining pool considered for training OLMo 3. (0 downloads, 0 stars, odc-by, pretraining, web-text, multi-source)
• Nemotron-Post-Training-Dataset-v2 — Multilingual Nemotron post-training data supporting Nemotron-Nano-9B-v2. (0 downloads, 0 stars, cc-by-4.0, sft, reinforcement-learning, multilingual)
• Aya Dataset — Cohere For AI's human-curated multilingual instruction dataset across 65+ languages. (0 downloads, 0 stars, apache-2.0, sft, multilingual, human-curated)
• AM-DeepSeek-R1-Distilled-1.4M — 1.4M verified general reasoning problems, ~0.9M distilled from DeepSeek-R1-671B. (0 downloads, 0 stars, cc-by-nc-4.0, reasoning, distilled, deepseek-r1)
• OpenR1-Math-220k — 220k verified math reasoning traces distilled from DeepSeek-R1 on NuminaMath 1.5. (0 downloads, 0 stars, apache-2.0, reasoning, math, distilled)
• FineWeb-Edu — Education-filtered subset of FineWeb web text for pretraining. (0 downloads, 0 stars, odc-by, pretraining, web-text, filtered)
• LMSYS-Chat-1M — 1M real user-model conversations from 25+ LLMs collected via Chatbot Arena. (0 downloads, 0 stars, custom (gated), conversations, real-user, research)
• Capybara — High-quality multi-turn synthetic conversations emphasizing reasoning and depth. (0 downloads, 0 stars, apache-2.0, sft, multi-turn, reasoning)
• Orca-AgentInstruct-1M-v1 — Microsoft's ~1M-example synthetic agentic-instruction dataset. (0 downloads, 0 stars, cdla-permissive-2.0, sft, agentic, synthetic)
• Deepseek-v4-pro-max-distill-1500x — Coding and math reasoning traces distilled from DeepSeek V4 Pro Max. (0 downloads, 0 stars, unspecified (source: DeepSeek V4 Pro Max, MIT + explicit distillation permission), reasoning, distilled, code)
• Databricks Dolly 15k — 15k human-written instruction/response pairs across 8 task categories. (0 downloads, 0 stars, cc-by-sa-3.0, sft, human-written)
• OpenHermes-2.5 — ~1M GPT-4-generated instruction pairs deduplicated to ShareGPT/ChatML format. (0 downloads, 0 stars, unknown, sft, instruction-tuning, gpt4-distilled)
• Magpie-Pro-300K (Filtered) — Filtered Magpie-Pro synthetic instructions self-synthesized from Llama-3-70B-Instruct. (0 downloads, 0 stars, llama3, sft, synthetic, filtered)
• qwen3-coder-480b-distill-mini — 9,543 cleaned code-reasoning samples distilled from Qwen3-Coder-480B-A35B-Instruct. (1 downloads, 3 stars, apache-2.0, code, reasoning, distilled)
• Nemotron-SFT-OpenCode-v1 — Agentic instruction-tuning data for the OpenCode CLI framework. (1 downloads, 1 stars, cc-by-4.0 (per Nemotron post-training family; not independently re-verified on this specific card), tool-calling, agentic, skills)
• Nemotron-SWE-v1 — 59k agent trajectories targeting SWE-Bench-style repo navigation and issue-fixing. (0 downloads, 0 stars, cc-by-4.0 (a few subsets are bsd-3-clause — see HF dataset viewer), tool-calling, agentic, code)
• When2Call — Teaches when NOT to call a tool: 15k SFT + 9k DPO preference pairs. (0 downloads, 0 stars, cc-by-4.0, tool-calling, agentic, preference)
• Nemotron-SFT-Agentic-v2 — ~1.2M synthetic tool-use trajectories seeded from UltraTool, ToolEyes, and more. (0 downloads, 0 stars, cc-by-4.0, tool-calling, agentic, multi-turn)
• Nemotron-Agentic-v1 — 335k synthetic multi-turn trajectories for interactive tool use. (0 downloads, 0 stars, cc-by-4.0, tool-calling, agentic, multi-turn)
• CodeFeedback Filtered Instruction — 156k execution-filtered code instructions — every solution passed its tests (Apache-2.0). (0 downloads, 0 stars, apache-2.0, coding, sft, execution-filtered)
• KIMI-K2.5-550000x — 550,000 high-depth reasoning traces distilled from Kimi K2.5. (0 downloads, 0 stars, unspecified (source: Kimi K2.5, Modified MIT + no distillation restriction), reasoning, distilled, kimi-k2.5)
• Hermes 3 Dataset — NousResearch's official post-training dataset behind the Hermes model line. (0 downloads, 1 stars, apache-2.0, sft, instruction-tuning, reasoning)
• SlimOrca — Deduplicated, cleaned subset of Orca with GPT-4-quality reasoning traces. (2 downloads, 2 stars, mit, sft, reasoning, deduplicated)

Popular tags

sft (36) · reasoning (25) · distilled (24) · agentic (21) · tool-calling (20) · coding (19) · open-harness (8) · math (8) · cot (6) · code (6) · nvidia (6) · deepseek-v4 (5) · non-commercial (5) · deepseek-r1 (5) · grpo (5) · instruction-tuning (5) · pretraining (4) · synthetic (4) · stem (4) · multilingual (4)

Recent activity

• Anonymous downloaded DeepSeek-V4-Pro 0813 Agentic (DS4)
• hf-mirror posted apigen-function-calling
• hf-mirror posted tool_calling
• hf-mirror posted xlam_hermes_validated
• hf-mirror posted tool-use-multiturn-reasoning
• hf-mirror posted tool-calls-singleturn
• hf-mirror posted RLVR-IFeval
• hf-mirror posted openchat_sharegpt4_dataset
• hf-mirror posted no_robots
• hf-mirror posted OpenOrca

For AI agents

This page is a server-rendered summary. The full machine-readable entry points:

• llms.txt (site index) — llms-full.txt
• OpenAPI contract for the public API at /api
• ARD capability manifest
• Live dataset JSON, model JSON