datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
dsv4-eval-artifacts
DeepSeek-V4-Flash-0731 — quantization measurements
Everything needed to reproduce, audit or extend the numbers published in
AtomicChat/DeepSeek-V4-Flash-0731-GGUF:
the reference logits, the evaluation corpus, the raw tool output for every quant we
measured, and the parsed results.
Every GGUF of this model that we could find on the Hub was measured here — ours,
unsloth's, bartowski's, ggml-org's, antirez's and others — on one machine, against one
reference, with one command.… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/dsv4-eval-artifacts.atomic-metrics-six-task-preferences
Six-task benchmark inputs
Seed 17. No demographic conditioning. Each task has shared train100.jsonl and test500.jsonl for Atomic Metrics, five judge variants, and learned baselines. Pair plans cover all 100 training rows once. Atomic Metrics extraction and BT/LR fitting use train100. Judges use the same test500. RM and WIMHF in the matched-data comparison use train100; rm_train_full is an explicitly separate expanded-data setting and must not be described as train100.… See the full description on the dataset page: https://huggingface.co/datasets/tintin1027/atomic-metrics-six-task-preferences.atomic-metrics-rm-splits
Atomic Metrics RM Task Splits
Preference-pair benchmark splits used by
Atomic Metrics. The release
contains four open-ended task families derived from public SHP, OASST1, and
OASST2 preference data.
Dataset structure
Each configuration contains 10,000 training pairs and 2,000 test pairs. Every
row has:
{
"sample_id": "source-specific stable ID",
"source_dataset": "shp | oasst1 | oasst2",
"category": "task configuration",
"split": "train | test"… See the full description on the dataset page: https://huggingface.co/datasets/tintin1027/atomic-metrics-rm-splits.OpenTopics-1.0-20K
OpenTopics-1.0-20K
What is this dataset?
OpenTopics-1.0-20K is a collection of 20,003 topic names spanning a wide variety of subjects, including physics, medicine, history, law, engineering, and the arts.
AtomixLabs built this dataset to help developers, researchers, and AI builders who need a large, organized list of topics. It works great for creating synthetic prompts, testing search systems, and training models to classify text.
What is inside… See the full description on the dataset page: https://huggingface.co/datasets/AtomixLabs/OpenTopics-1.0-20K.atomic-formal-reasoning-complex
Atomic Formal Reasoning — Complex Numbers
Overview
This dataset contains high-quality Lean 4 formal proofs of complex number theorems, written in an explicit pedagogical calc-chain style. Each proof is fully verified, step-by-step, with no opaque tactics (simp, ring, omega are avoided). Every reasoning step is named and justified.
This is process supervision data — not just final answers. Each entry exposes the full reasoning chain, making it ideal for training models… See the full description on the dataset page: https://huggingface.co/datasets/7rouz/atomic-formal-reasoning-complex.nemotron-cc-atomic-simplification-gemma4-31b
nemotron-cc atomic-statement simplification (Gemma 4 31B-it)
2,000,000 records: source text from
nvidia/nemotron-cc-v2.1 (High-Quality-Synthetic
split) rewritten by google/gemma-4-31B-it into a sequence of atomic, Subject-Verb-Object
statements.
Generated with vLLM 0.22.1 in-process batch inference (see src/generate/run.py in the
producing repo), TP=4, max_model_len=16384, max_tokens=8192, prompts filtered to
<=8192 templated tokens.
Fields
id: original… See the full description on the dataset page: https://huggingface.co/datasets/rpisano/nemotron-cc-atomic-simplification-gemma4-31b.sydney-training-data
Sydney 训练集
四份来源分开存放，不混在一个文件里。
发布的聊天权重（Atonelia/Qwen3.5-Sydney-9B / -think 以及对应 GGUF）用的是这些子集洗完、抽样拼起来之后的训练 jsonl，不是直接拿某一份原文训的。
01 原截图重建
01_screenshot_original/conversations.jsonl
早期 Bing Chat / Sydney（约 2023 年 2–4 月）公开截图重建的对话。660 条，原文以英文为主，带截图出处。
这是最初拿来做训练集的底。后面的中文版、合成版、CoT 都不是这份文件本身。
02 llama-sydney 虚拟对话
02_llama_sydney_synthetic/llama_sydney_en.jsonl
用 Llama-Sydney 生成的英文虚拟对话。1462 条（同一条 user 可能有 2 次采样）。字段是生成记录：id / user / assistant 等，还不是最终训练格式。… See the full description on the dataset page: https://huggingface.co/datasets/Atonelia/sydney-training-data.ATOM_regen_seeklight_kimi_mtp
ATOM regen: seeklight kimi-mtp responses by Kimi-K3
用 Kimi-K3 对
lightseekorg/kimi-mtp-dataset
的 prompt 重新生成了一遍回答，供 off-policy 投机解码蒸馏（SDDD）使用。
原始 pipeline 每轮都要用 teacher 重新解码一次（Phase A1）。把回答预生成并缓存下来，
A1 整个消失，之后每一轮训练都直接复用，代价从「每轮一次」变成「一共一次」。
数据
450,625 行，每行一段对话：
{"conversations": [
{"role": "user", "content": "..."},
{"role": "assistant", "reasoning_content": "...", "content": "..."}
]}
reasoning_content 是 K3 的 thinking 内容，和 content 分开存。
多轮对话保留了历史轮次里完整的 assistant… See the full description on the dataset page: https://huggingface.co/datasets/slippedJim/ATOM_regen_seeklight_kimi_mtp.atos-eviden-chatATOM-regen-nine-category-prompt
ATOM regen — Nemotron nine-category pool + aya, answered by Kimi-K3
Every prompt from the nine nvidia/Nemotron-Post-Training-Dataset-v2 splits plus
CohereLabs/aya_dataset re-answered by Kimi-K3, then verified token-by-token.
An off-policy dataset you can reuse without re-running decode.
5,030,694 rows · 11.7 B generated tokens
Provenance
Teacher
moonshotai/Kimi-K3 (official repo)
Weights revision
a590ce090cb049c93a33dfe8c208ec652aa20503
Engine… See the full description on the dataset page: https://huggingface.co/datasets/slippedJim/ATOM-regen-nine-category-prompt.ATOM_regen_nv_OpenCodeInstruct
ATOM regen: OpenCodeInstruct responses by Kimi-K3
用 Kimi-K3 对
nvidia/OpenCodeInstruct
的 prompt 重新生成了一遍回答，供 off-policy 投机解码蒸馏（SDDD）使用。
窗口 32768。
为什么是代码，为什么是这个数据集
训练一个 5 层 DSpark draft 的过程里，代码一直是最薄的一片：主语料去重之后
code 只剩 23,083 行唯一 prompt（17.5 万行里平均一题 5.6 个回答，17.9% 是物理
上限），占整个语料的 **0.43%**。对应的后果是 SWE-bench Pro 成为第二弱的
benchmark —— 只有同机官方 draft 的 **84.9%**。
把代码占比从 0.43% 提到 **20%**（主要靠这份数据）之后，SWE-bench Pro 到了
**100.6%**，是 13 套 benchmark 里相对提升最大的一套。
这份数据还顺带解决另一件事：长 prompt… See the full description on the dataset page: https://huggingface.co/datasets/slippedJim/ATOM_regen_nv_OpenCodeInstruct.