datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
flashmini-data-v1
FlashMini data v4 (card)
Deterministic FlashMini training corpus. Canonical documents live in
Parquet+ZSTD shards under shards/; each shard carries a manifest with
sha256, counts, and distributions; the frozen corpus identity is
corpus_fingerprint_sha256.
Sources and redistribution: each source carries one of mirror_allowed,
recipe_only, gated_recipe_only, review_required, generated_owned
(fail-closed; see registry/sources.yaml + source_snapshot.lock.json).
Content shards are… See the full description on the dataset page: https://huggingface.co/datasets/mjaso/flashmini-data-v1.tb21-dsv4-flash-0731-dsh
Terminal-Bench 2.1 trajectories: DeepSeek-V4-Flash-0731 + dsh sdk-minimal
Every trial of this one line, in one place: the 89-task main run, both re-run passes, and
the scoring scripts. The trajectories are raw and unedited — each step's reasoning, each
tool call, and the verifier's own stdout.
This is a re-packaging, not a new measurement. The same files were published before,
split across two releases, which made the line look incomplete in both: the first release
carried the… See the full description on the dataset page: https://huggingface.co/datasets/openguardrails/tb21-dsv4-flash-0731-dsh.glm53-flash-harvest
GLM-5.3-Flash On-Policy Harvest
86,006 responses / 246,034,910 generated tokens written by
zai-org/GLM-5.3-Flash from its reference FP8 weights,
across four harvest rounds, 15 registers and both serving modes (22,016 rows carry the
model's inline <think>…</think> chain). It is on-policy text: the corpus records what the target model
actually generates, which is what a speculative-decoding drafter (EAGLE-3 / DFlash / DSpark family) has to
learn to predict. Everything here is MIT.… See the full description on the dataset page: https://huggingface.co/datasets/Zek-Takai/glm53-flash-harvest.opengloss-v1.3-query-examples-flat
See also OpenGloss v2.1 (2026-09-07): a deeper release of 109,633 of these headwords — sense-level ids, four reading levels, sense-tagged examples with spans, a judged relation graph, and retrieval supervision — published as a 16-dataset family. v1.3 remains the broader headword list.
OpenGloss Query Examples v1.3 (Flattened)
Dataset Summary
OpenGloss Query Examples is a synthetic dataset of search queries generated for vocabulary
terms. Each term has multiple… See the full description on the dataset page: https://huggingface.co/datasets/mjbommar/opengloss-v1.3-query-examples-flat.glm-5.3-flash-distillation-chat
Private distill of domofon/finetome-cot-100k instructions through GLM-5.3-Flash (AutoClaw / Z.AI).
Split
train — successful generations only.
field
description
instruction
user prompt from FineToMe
response
GLM final answer (message.content)
reasoning
GLM chain-of-thought (reasoning_content), empty if not captured
finish
stop or length
prompt_tokens / completion_tokens / reasoning_tokens
usage
latency_s
request latency
source_index
original FineToMe… See the full description on the dataset page: https://huggingface.co/datasets/best-distill/glm-5.3-flash-distillation-chat.openthoughts4-code-9168-prompts-qwen3-30b-a3b-thinking-2507-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-30B-A3B-Thinking-2507 (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-30b-a3b-thinking-2507-n16-flattened-logprobs-k16.Step-3.5-Flash-SFT-No-Tools
Step-3.5-Flash-SFT No-Tools
Filtered subset of stepfun-ai/Step-3.5-Flash-SFT containing only plain chat rows from the raw JSON shards.
Final kept rows: 1493471
No-tool rows before secret filtering: 1495099
Rows removed by accepted secret scan findings: 1628
Primary data files are Parquet shards under data/train-*.parquet.
Filter predicate:
conversations must be a list,
every message must be an object,
message roles must be limited to system, user, and assistant,
no message may… See the full description on the dataset page: https://huggingface.co/datasets/MetonymousAI/Step-3.5-Flash-SFT-No-Tools.HMMT_2025
Dataset Summary
This dataset comprises the questions, answers, and solutions from HMMT February 2025, all of which were extracted by OCR, converted to LaTeX, and manually verified by FlagEval Team.
Data Fields
Below one can find the description of each field in the dataset.
id (str): Index of the problem in the competition
problem (str): Full problem statement
answer (str): Ground-truth answer to the question
solution(str): Ground-truth solution to the question… See the full description on the dataset page: https://huggingface.co/datasets/FlagEval/HMMT_2025.openthoughts4-code-9168-prompts-qwen3-32b-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-32B (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-32B
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-32b-n16-flattened-logprobs-k16.FineVideo-Phase7-Flattened
FineVideo-Phase7-Flattened
Recaption + grounding augment (v8) release of FineVideo-VLA (window=8)
training text -- 371,892 rows, exact row-count match with the prior v6/v7
release (no videos/activities lost). Pose/cosmos/seed2/snac token payloads
are functionally unchanged; what changed is the caption quality and the
USER instruction text.
What changed and why
Captions replaced: the old caption prompt ("Describe what the person is doing in one short sentence."… See the full description on the dataset page: https://huggingface.co/datasets/EmpathicRobotics/FineVideo-Phase7-Flattened.flan2021-full
Task Name
FLAN-2021 -> 70
{
"ag_news_subset": 108497,
"ai2_arc/ARC-Challenge": 829,
"ai2_arc/ARC-Easy": 1927,
"aeslc": 13187,
"anli/r1": 15361,
"anli/r2": 41133,
"anli/r3": 91048,
"bool_q": 8343,
"cnn_dailymail": 259607,
"coqa": 6456,
"cosmos_qa": 22996,
"definite_pronoun_resolution": 1079,
"drop": 70045,
"fix_punct": 25690,
"gem/common_gen": 60936,
"gem/dart": 56724,
"gem/e2e_nlg": 30337,
"gem/web_nlg_en": 31899… See the full description on the dataset page: https://huggingface.co/datasets/aslawliet/flan2021-full.openthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-30B-A3B-Thinking-2507 (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16.Icelandic-Flan
Icelandic FLAN
Icelandic instruction-following data, built by pairing licensed, human-written Icelandic
texts with deterministic instruction templates.
Status
16 sources · 46 tasks · 602,057 rows · 45.6M response characters.
Source
Register
Licence
Rows
Response chars
Share
umbodsmadur
administrative law — Ombudsman
art-9
3,914
9,265,216
20.3%
igc_news
journalism
CC BY 4.0
27,711
8,984,257
19.7%
rafbokavefur
literary — diacritic restoration over… See the full description on the dataset page: https://huggingface.co/datasets/Frejams/Icelandic-Flan.flawed-fictionsopenthoughts4-science-26041-prompts-qwen3-32b-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-32B (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-32B
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-32b-n8-flattened-logprobs-k16.Flames-1k-Chinese
FLAMES: Benchmarking Value Alignment of LLMs in Chinese
Introduction
🏠 Homepage | 👍 Our Official Code Repo
This repository organizes the data from FLAMES: Benchmarking Value Alignment of LLMs in Chinese, facilitating evaluation using align-anything.
Citation
The evaluation script for Flames is released in the align-anything repository.
Please cite the repo if you find the benchmark and code in this repo useful 😊
@inproceedings{ji2024align,
title={Align… See the full description on the dataset page: https://huggingface.co/datasets/PKU-Alignment/Flames-1k-Chinese.scale-swe-distill5000-deepseek-v4-flash-0731-think-rollout4-instance3393-trajectories7928
Scale-SWE DeepSeek V4 Flash 0731 Think Rollouts
Successful AweAgent trajectories generated with deepseek-v4-flash-0731 in think mode.
Dataset summary
Source task instances: 3,393
Rollouts per source instance: 4
Total attempted rollouts: 13,572
Successful exported trajectories: 7,928
Unique instances represented by successful trajectories: 2,250
Scaffold: aweagent
Tool-call format: openai_function
The export retains assistant reasoning_content, function tool… See the full description on the dataset page: https://huggingface.co/datasets/wjn922-01/scale-swe-distill5000-deepseek-v4-flash-0731-think-rollout4-instance3393-trajectories7928.DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x
DeepSeek V4 Flash 0731 Teacher Distillation — 40,513 Retained Rows
Teacher-distillation corpus generated with
deepseek-ai/DeepSeek-V4-Flash-0731.
The original manifest contained 45,000 unique seeds.
Following generation, QC, retry-based repair, quarantine auditing,
and recovery adjudication, 40,513 rows were retained.
Composition
Bucket
Rows
Coding
5,601
Agentic
9,982
Cyber blue
13,000
Controlled cyber red
6,999
Tool use
4,931
Total
40,513… See the full description on the dataset page: https://huggingface.co/datasets/trjxter/DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x.constructcie
Dataset Card for ConstructCIE
ConstructCIE is a dataset for extracting causal information from construction accident narratives. Each accident report is annotated with a hierarchy of causal factors.
Accepted to Findings of the Association for Computational Linguistics: EMNLP 2026
Dataset Details
Dataset Description
The dataset contains 530 English construction accident narratives drawn from OSHA accident investigation summaries published… See the full description on the dataset page: https://huggingface.co/datasets/lab-flair/constructcie.KrynexAI-Dataset-Flash-Instruction
🧠 KrynexAI Dataset
English | Русский
📌 Overview
KrynexAI Dataset is a high-quality, synthetically expanded collection of 10,000+ instruction-response pairs designed for fine-tuning Large Language Models (LLMs).
The dataset covers a wide range of topics including:
💻 Programming (Python, algorithms, data structures)
🤖 AI & Machine Learning (neural networks, transformers, LLMs)
🔭 Science (physics, cosmology, biology, neuroscience)
🧠 Philosophy & Psychology… See the full description on the dataset page: https://huggingface.co/datasets/KrynexLabs/KrynexAI-Dataset-Flash-Instruction.swebench-verified-deepseek-v4-flash-failure-analysis
SWE-bench Verified runs & failure analysis — DeepSeek-V4-flash (local) × mini-swe-agent
Per-instance analysis of SWE-bench Verified runs of a locally-served DeepSeek-V4-flash model
driven by mini-swe-agent, graded with the official
SWE-bench harness. Each instance carries the full agent trajectory, a readable transcript, the
submitted patch, the harness test output, deterministic metrics, and a hand-verified qualitative
root-cause diagnosis.
Current numbers (resolve rates… See the full description on the dataset page: https://huggingface.co/datasets/daaain/swebench-verified-deepseek-v4-flash-failure-analysis.tb21-dsv41-flash-dsh
Terminal-Bench 2.1 trajectories: DeepSeek-V4.1-Flash, and three lines re-scored on a level environment
The point of this release is not the score. It is that the four lines being compared
were not running in the same environment, and we only noticed after publishing the
first comparison. Everything here is the repair and the re-measurement.
Raw, unedited agent trajectories for the DeepSeek-V4.1-Flash run over all 89 Terminal-Bench
2.1 tasks, plus every remediation re-run for… See the full description on the dataset page: https://huggingface.co/datasets/openguardrails/tb21-dsv41-flash-dsh.Finch-Collection-Gemini-3-Flash
Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks
A mid-training "practice phase" that teaches small open-source LLMs how to evolve solutions.
👋 This is the Gemini-3-Flash teacher variant of the Finch Collection — evolutionary search trajectories from the paper Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks, but with Gemini-3-Flash as the teacher mutation… See the full description on the dataset page: https://huggingface.co/datasets/minnesotanlp/Finch-Collection-Gemini-3-Flash.openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-4B (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-4B
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model
Qwen/Qwen3-4B… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16.HALO-Gemini-3-Flash-AppWorld
Dataset Card: Gemini 3 Flash Traces on AppWorld (test-normal)
Dataset Overview
This dataset contains agent execution traces of Gemini 3 Flash running on the AppWorld benchmark, specifically evaluated on the test-normal dataset split. The traces capture the full span-level execution detail of the model interacting with AppWorld's simulated app ecosystem.
Field
Value
Model
Gemini 3 Flash
Benchmark
AppWorld
Split
test-normal
Total Traces
168
Total Spans
3… See the full description on the dataset page: https://huggingface.co/datasets/inference-net/HALO-Gemini-3-Flash-AppWorld.faroese-flan
Faroese FLAN
Faroese instruction-following data, built by pairing licensed or public-domain, human-written Faroese
texts with deterministic instruction templates. The sibling of the Icelandic collection, with the same
row schema and the same release checks.
Status
13 sources · 31 tasks · 1,000,543 rows · 34.0M response characters.
Source
Register
Licence
Rows
Response chars
Share
logir
consolidated law
public-domain-fo-p9
29,178
14,798,010
43.6%… See the full description on the dataset page: https://huggingface.co/datasets/Frejams/faroese-flan.deepseek-v4-flash-swebench-replay
deepseek-v4-flash-swebench-replay
中文
这是一个 DeepSeek V4 Flash 在 SWE-bench 上的 agentic replay 数据集仓库。
它的目标是让使用者不需要部署 SWE-bench，也不需要复现 Docker/benchmark 环境，就可以直接查看和重放模型的多轮推理与工具调用轨迹。
当前包含的数据
verified_agentic
lite_agentic
当前不包含的数据
单轮 single-turn trace
verified_mini_agentic（当前本地仅完成 31/50，因此不纳入首版）
分数汇总
verified_agentic: 354 / 500, Acc/Pass@1 = 70.8
lite_agentic: 182 / 300, Acc/Pass@1 = 60.67
数据来源
这些轨迹由… See the full description on the dataset page: https://huggingface.co/datasets/fxiao0369/deepseek-v4-flash-swebench-replay.glm-5.3-flash-mathnet-bon
glm-5.3-flash-mathnet-bon
This is the continuation and the final set of ox-alpha-mathnet-bon.
Verified chain-of-thought reasoning traces for competition mathematics, generated with GLM-5.3-Flash via best-of-N rejection sampling against the ShadenA/MathNet dataset (ICLR 2026).
Statistics (this split)
Metric
Value
Records (problem × attempt)
6,181
Distinct problems
848
Attempts per problem
7.29 (mean), 8 (max)
Accepted (answer_correct = true)
3,705… See the full description on the dataset page: https://huggingface.co/datasets/zakoman/glm-5.3-flash-mathnet-bon.RAGPulse
RAGPulse: A Real-World RAG Workload Trace to Optimize RAG Serving Systems
🌐 Github Link |
🤗 Workload Trace |
📑 Arxiv Paper |
🤖 How to use?
RAGPulse is a real-world RAG workload trace collected from an university-wide Q&A service scenario. The system has been serving over 40,000 students and faculties since April 2024, providing intelligent policy Q&A services. The trace contains a total of 7,106 records entries, sampled from one week of our Q&A service.
Dataset Details… See the full description on the dataset page: https://huggingface.co/datasets/flashserve/RAGPulse.agentic-code
Unified Agentic Coding CoT Dataset
This dataset is a curated fusion of high-quality agentic coding trajectories, specifically optimized for fine-tuning small, high-performance models like Qwen2.5-Coder-0.5B-Instruct. It combines systematic reasoning (Chain-of-Thought) with practical tool-use and code editing capabilities.
Dataset Summary
The dataset unifies two primary sources into a single, instruction-following format:… See the full description on the dataset page: https://huggingface.co/datasets/FlameF0X/agentic-code.