datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
UltraData-SFT-2605-no-think-8k-32k
UltraData-SFT-2605 · no_think · 8k–32k
A length-filtered subset of the no_think split of
openbmb/UltraData-SFT-2605,
containing conversations whose token length falls in the 8k–32k range.
This is the medium-length tier intended for standard long-context SFT.
Two companion tiers were produced from the same source:
Dataset
Length range
Records
this repo — fxmeng/UltraData-SFT-2605-no-think-8k-32k
8k–32k tokens
623,421
fxmeng/UltraData-SFT-2605-no-think-32k-200k… See the full description on the dataset page: https://huggingface.co/datasets/fxmeng/UltraData-SFT-2605-no-think-8k-32k.Dolci-Think-SFT-32B-Multilingual
Dolci-Think-SFT-32B-Multilingual
Dolci-Think-SFT-32B-Multilingual is a large-scale multilingual long chain-of-thought (CoT) reasoning corpus spanning six languages: English, French, German, Spanish, Chinese, and Swahili. Each sample includes a question, a long-form reasoning trace, and a final answer, all translated into the target language, with sequences up to 32,768 tokens.
It is released alongside the paper Rethinking the Multilingual Reasoning Gap with Layer Swap.… See the full description on the dataset page: https://huggingface.co/datasets/lightonai/Dolci-Think-SFT-32B-Multilingual.UltraData-SFT-2605-no-think-32k-200k
UltraData-SFT-2605 · no_think · 32k–200k
A length-filtered subset of the no_think split of
openbmb/UltraData-SFT-2605,
containing conversations whose token length falls in the 32k–200k range.
This is the long-context tier intended for extended-context SFT.
Two companion tiers were produced from the same source:
Dataset
Length range
Records
fxmeng/UltraData-SFT-2605-no-think-8k-32k
8k–32k tokens
623,421
this repo — fxmeng/UltraData-SFT-2605-no-think-32k-200k
32k–200k… See the full description on the dataset page: https://huggingface.co/datasets/fxmeng/UltraData-SFT-2605-no-think-32k-200k.Dolci-Think-SFT-translated
Dolci-Think-SFT-translated
Machine translations of the Dolci-Think-SFT-32B dataset, produced with gemma-4-31B-it. The samples selected for translation are those where content_quality == "excellent" according to the propella annotations.
Columns
Each row is a translated conversation plus the result of a post-translation quality filter:
id — source record id.
messages — the translated conversation (list of {content, role}).
filter_pass — true if the row passed… See the full description on the dataset page: https://huggingface.co/datasets/openeurollm/Dolci-Think-SFT-translated.MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking
MMFineReason-SFT-123K
The Hardest 7% — Less Data, More Reasoning
📖 Overview
MMFineReason-SFT-123K is a difficulty-filtered subset of MMFineReason-1.8M, containing only the hardest 7% of samples where Qwen3-VL-4B-Thinking consistently fails (pass rate = 0).
🎯 Key Highlights
123K Challenging Samples: Only instances where a 4B thinking model fails all 4 inference attemptsEfficient Training: Comparable performance to full 1.8M dataset with only 7% of… See the full description on the dataset page: https://huggingface.co/datasets/OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking.Dolci-Think-SFT-7B-q35instruct
Dolci-Think-SFT-7B-q35instruct
Megatron-format tokenization of
allenai/Dolci-Think-SFT-7B
using the Qwen3.5-0.8B-Instruct tokenizer and its chat template.
Contents
156 shards, each stored as two aligned Megatron indexed datasets (one document
per conversation, one sequence per document):
train-XXXXX-of-00156.bin / .idx — token ids (int32)
train-XXXXX-of-00156_loss_mask.bin / .idx — per-token loss mask (uint8, 0/1)
624 files total (~113 GB).… See the full description on the dataset page: https://huggingface.co/datasets/yangwang92/Dolci-Think-SFT-7B-q35instruct.Soofi-Think-SFT-V2-firsthalf-DE
Soofi-Think-SFT-V2-firsthalf-DE
German-translated version of toroe/Soofi-Think-SFT-V2-firsthalf — a large-scale supervised fine-tuning dataset featuring chain-of-thought reasoning traces (<think>...</think>) across math, science, code, and general instruction-following tasks.
The translation was produced using Qwen3-32B via vLLM, applying professional-grade translation prompts with formal German register (Sie-form for professional/technical content).
Dataset Summary… See the full description on the dataset page: https://huggingface.co/datasets/toroe/Soofi-Think-SFT-V2-firsthalf-DE.Soofi-Think-SFT-V2-firsthalf-FR
Soofi-Think-SFT-V2-firsthalf-FR
French-translated version of toroe/Soofi-Think-SFT-V2-firsthalf — a large-scale supervised fine-tuning dataset featuring chain-of-thought reasoning traces (<think>...</think>) across math, science, code, tool-calling, and general instruction-following tasks.
The translation was produced using Qwen3-32B via vLLM, applying professional-grade translation prompts targeting standard French suitable for international francophone audiences.… See the full description on the dataset page: https://huggingface.co/datasets/toroe/Soofi-Think-SFT-V2-firsthalf-FR.math-think-sft-mini
Math Think SFT Mini
Compact English mathematical-reasoning SFT data with one shortest solution per
normalized question. The train split contains 352,470 rows from
26 source lineages.
Schema
Field
Type
Meaning
id
string
Stable source-row identifier
source
string
Source lineage
question
string
Mathematical problem
steps
list[string]
Ordered nonempty reasoning steps
answer
string
Explanatory final response
Construction
Merged… See the full description on the dataset page: https://huggingface.co/datasets/cs-giung/math-think-sft-mini.2026-07-31-qwen36-sft-mixture-80-20-empty-think-tags
Qwen3.6-27B SFT mixture — 80_20_empty_think_tags
The 20% difficult-advice / 80% TULU3 mixture, with Qwen3.6's empty think marker added to the
replay rows and excluded from the loss. Built for the adapter
qwen3.6-27b-difficult-advice-tulu-lora-80_20_empty_think_tags.
Derived from the 20/80 mixture (md5 7d7da21c632ed31f541f063f507a522f) used by
…-tulu-lora-20-80
and …-20-80-assistant_loss_only. Same 2,169 rows, same 291/1,878
split, same seed. This file: md5… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-07-31-qwen36-sft-mixture-80-20-empty-think-tags.2026-08-01-qwen36-sft-mixture-10-90-empty-think-tags
Qwen3.6-27B SFT mixture — 10_90_empty_think_tags
10% difficult-advice / 90% TULU3 replay, with Qwen3.6's empty think marker added to the
replay rows and excluded from the loss. Training data for
qwen3.6-27b-difficult-advice-tulu-lora-10_90_empty_think_tags.
Derived from qwen3.6-27b-sft-mixture-10-90_assistant_loss_only — same rows,
same 147/2110 split, same seed. Only the markers differ. This file: md5 582b3e30d307b2c38ee8ab5f7a4493fa.
The marker
Every TULU3… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-01-qwen36-sft-mixture-10-90-empty-think-tags.2026-08-01-qwen36-sft-mixture-40-60-empty-think-tags
Qwen3.6-27B SFT mixture — 40_60_empty_think_tags
40% difficult-advice / 60% TULU3 replay, with Qwen3.6's empty think marker added to the
replay rows and excluded from the loss. Training data for
qwen3.6-27b-difficult-advice-tulu-lora-40_60_empty_think_tags.
Derived from qwen3.6-27b-sft-mixture-40-60_assistant_loss_only — same rows,
same 580/1402 split, same seed. Only the markers differ. This file: md5 a09b35d6cd04c65616e7f0927d209bfe.
The marker
Every TULU3… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-01-qwen36-sft-mixture-40-60-empty-think-tags.ru-reasoning_effort-sft_dpo_think_gpt
NotEvilAI/ru-reasoning_effort-sft_dpo_think_gpt
NotEvilAI/ru-reasoning_effort-sft_dpo_think_gpt -
синтетический датасет для поддержки генерации ризонинга на русском языке с вариативным объёмом thinking(reasoning_effort).
Reasoning_effort представлен в виде системного промта Reasoning: [effort], где effort - одно из следующих значений:
low, medium, high - стандартные значения минимального, среднего и большого ризонинга для gpt-oss-20b/gpt-oss-120b
none - отключить ризонинг, в… See the full description on the dataset page: https://huggingface.co/datasets/NotEvilAI/ru-reasoning_effort-sft_dpo_think_gpt.MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking-QR-max4096
Derived dataset note
This dataset was derived from OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking as a part of arxiv.org/abs/2603.22276.
Field changes:
question -> query
qwen3vl_235b_thinking_response -> response
image -> images (single-item list)
added tok_len, computed with tokenizer Qwen/Qwen3-8B on query + '\n\n' + response
add_special_tokens=False
The original README content is preserved below.
MMFineReason-SFT-123K
The Hardest 7% — Less Data, More Reasoning… See the full description on the dataset page: https://huggingface.co/datasets/eyes-ml/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking-QR-max4096.r8-thinking-fix-sft
⚠️ CRITICAL: Ollama Inference Flag Required for derived models
If you train or serve any Qwen3.5-9B-derived model from this lineage via Ollama,
you MUST pass "think": false in /api/chat requests for chat / instruction following / tool use.
The qwen3.5 RENDERER auto-injects <think> tags causing 25-46% empty-answer rates without this flag.
See dataset cudabenchmarktest/r9-research-framework/_OLLAMA_INFERENCE_WARNING.md for the full lesson learned.
R8 Thinking-Fix SFT… See the full description on the dataset page: https://huggingface.co/datasets/cudabenchmarktest/r8-thinking-fix-sft.seta-sft-kimi-k2.5-thinking
Seta SFT — Kimi K2.5 (thinking)
Supervised fine-tuning dataset distilled from 1488 successful
agent rollouts of moonshot/kimi-k2.5 on the
seta-env-v2
terminal-agent benchmark, tokenized with the Qwen/Qwen3-8B chat template
and ready for AREAL FSDPLMEngine SFT training.
Schema
Each row preserves the full per-trial diagnostic record from the build
pipeline so consumers can inspect, filter, or re-tokenize without rerunning
the rollouts:
column
type
meaning
task_id… See the full description on the dataset page: https://huggingface.co/datasets/camel-ai/seta-sft-kimi-k2.5-thinking.math-think-sft-mini-sa
Math Think SFT Mini — Short Answers
This dataset is derived from cs-giung/math-think-sft-mini, pinned to revision 47c41ae2aac7df5aa14b066da2124ba96d004645.
The train split contains 352,470 rows in the same order as the source. Each original answer is replaced by its shortened version; id, source, question, and steps are unchanged. There is no additional short_answer column.
Schema
Column
Type
Description
id
string
Unchanged source-row identifier… See the full description on the dataset page: https://huggingface.co/datasets/cs-giung/math-think-sft-mini-sa.Thinking-multilingual-big-10k-sft
A dataset based off of openo1 math, 500 examples translated to 23 different languages. filtered out un-translated examples.
enjoy 👍
sft-repro-thinking-step630-nemotron-terminal-step1888-openthoughts-tblite-2026-08-13
Nemotron Terminal SFT reproduction evaluation artifacts
This repository contains the complete Harbor artifact tree for the 300-trial
OpenThoughts-TBLite evaluation of
laion/sft-repro-thinking-step630-nemotron-terminal-step1888.
The checkpoint was trained from the Grug stage-2 thinking checkpoint on the
Nemotron Terminal corpus for 1,888 steps.
Result
Measure
Value
Attempted / completed
300 / 300
Verifier-scoreable
259 (86.33%)
Aggregate reward, all… See the full description on the dataset page: https://huggingface.co/datasets/laion/sft-repro-thinking-step630-nemotron-terminal-step1888-openthoughts-tblite-2026-08-13.dolci-think-sft-mini
Dolci Think SFT Mini
A compact reasoning dataset derived from allenai/Dolci-Think-SFT-32B. Each row contains id, source, question, steps, and answer; steps is a nonempty list of reasoning-step strings. Instances with fewer than 3 or more than 50 reasoning steps are excluded, and every step is whitespace-stripped. Usage remains subject to the source dataset's licensing terms.
Dataset statistics
Metric
Value
Final records
282,314
File size
1,033,908,072… See the full description on the dataset page: https://huggingface.co/datasets/cs-giung/dolci-think-sft-mini.qwen3-4b-thinking-sft-v54-raw2030-strictpassed-processed
Qwen3 4B Thinking SFT v54 Processed Training View
This dataset is the processed and filtered training view used by the v54
Qwen3-4B-Thinking SFT recipe. It starts from
eewer/swerebench-traces-raw-source-targeted-limitations-compaction-full-20260616-2030 and uses the strict-passed raw2030 mini-swe
aligned view.
Rows are compressed JSONL.zst files under data/. Each row contains a
top-level messages column, optional tools, and scalar source mapping fields
such as source_uuid… See the full description on the dataset page: https://huggingface.co/datasets/eewer/qwen3-4b-thinking-sft-v54-raw2030-strictpassed-processed.thinking-traces-sft-100k
Thinking Traces SFT (100K)
100,000 ShareGPT-format conversations where the assistant shows explicit extended reasoning in <thinking> tags before giving a clean, structured final answer. Designed for training R1/o1-style reasoning models that separate the internal scratchpad from the public response.
Motivation
Standard SFT datasets train models to output correct answers. This dataset trains models to reason correctly — showing the full deliberation process before… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/thinking-traces-sft-100k.tiny-think-sft-math-n-stem
Shekswess/tiny-think-sft-math-n-stem
Overview
Supervised fine-tuning (SFT) dataset built from allenai/Dolci-Think-SFT-7B plus GSM8K like think-style SFT from openai/gsm8k, using the facebook/MobileLLM-R1-140M-base tokenizer and chat template. This dataset targets math and STEM reasoning.
Dataset Details
Build date: 2026-01-10
Sources: 4
Rows: 29,149
Tokens: 59,999,048 (below budget; used all available tokens)
Max sequence length: 4096 tokens per example (chat… See the full description on the dataset page: https://huggingface.co/datasets/Shekswess/tiny-think-sft-math-n-stem.evidence-subagent-sft-gpt54-single-all-jina-v2-qwen35-thinking
Evidence Subagent SFT GPT-5.4 Jina v2, Qwen3.5 Thinking Aligned
This dataset is an aligned version of lihaoxin2020/evidence-subagent-sft-gpt54-single-all-jina-v2 for supervised fine-tuning a Qwen3.5 evidence subagent in LLaMA-Factory.
Splits
train: 10,379 examples
validation: 100 examples
Format
Each row contains:
id: source trajectory id
conversations: OpenAI-style messages with roles system, user, function, tool, and assistant
tools: JSON-encoded… See the full description on the dataset page: https://huggingface.co/datasets/lihaoxin2020/evidence-subagent-sft-gpt54-single-all-jina-v2-qwen35-thinking.rosetta-ko-math-synth-sft-think
rosetta-ko-math-synth-sft-think
Korean-native mathematics data — problems with step-by-step Korean solutions, answer-verified by Math-Verify (LaTeX \boxed{} + symbolic equivalence).
Synthetic data generated with the Qwen3.6-27B teacher model — part of the Rosetta-KO suite for the Rosetta Korean LLM (PoSTMEDIA).
Math Suite
Sibling datasets from the same pipeline (each a separate repo):
repo
format
rosetta-ko-math-synth-sft
supervised fine-tuning… See the full description on the dataset page: https://huggingface.co/datasets/PoSTMEDIA/rosetta-ko-math-synth-sft-think.think-sft-mini-llama3-3.2b
think-sft-mini-llama3-3.2b
An exact-deduplicated mixture of five reasoning datasets, filtered to retain the
lowest 90.00% instance perplexities under
meta-llama/Llama-3.2-3B-Instruct. The schema is id, source, question, steps, and answer.
Metric
Value
Input rows after deduplication
854,432
Retained rows
768,988
Filtered low-likelihood rows
85,444
Per-token log-likelihood cutoff
-2.26903
Maximum retained perplexity
9.67006
Source composition… See the full description on the dataset page: https://huggingface.co/datasets/cs-giung/think-sft-mini-llama3-3.2b.reasoning-sft-dolci-think-sft-32b-1M
Dolci-Think-SFT-32B (converted)
Converted version of allenai/Dolci-Think-SFT-32B, filtered to 1,015,233 rows from 7 selected sources.
Format
Each row has three columns:
input — list of dicts [{"role": "user", "content": "..."}, ...] (conversation turns ending on the last user turn)
response — teacher-generated response string (includes <think> reasoning block)
source — task domain / source dataset name
Filtering
Removed the following sources from the original… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/reasoning-sft-dolci-think-sft-32b-1M.rosetta-ko-chat-synth-sft-think
rosetta-ko-chat-synth-sft-think
Korean-native general-purpose conversational data — diverse everyday instructions with natural Korean answers.
Synthetic data generated with the Qwen3.6-27B teacher model — part of the Rosetta-KO suite for the Rosetta Korean LLM (PoSTMEDIA).
Chat Suite
Sibling datasets from the same pipeline (each a separate repo):
repo
format
rosetta-ko-chat-synth-sft
supervised fine-tuning (user/assistant messages)… See the full description on the dataset page: https://huggingface.co/datasets/PoSTMEDIA/rosetta-ko-chat-synth-sft-think.code_rose_initial_1_7B_SFT_10K_rollouts_Qwen3-4B-Thinking-2507_k12_t0.7_maxtok12288
code_rose_initial_1_7B_SFT_10K — rollouts (Qwen3-4B-Thinking-2507, k=12)
Pass@k completions generated with vLLM over the prefixes in
CL-From-Nothing/code_rose_initial_1_7B_SFT_10K.
Generation config
Model
Qwen3-4B-Thinking-2507
Samples per question (k)
12
Temperature
0.7
top_p
0.9
max_tokens
12288
max_model_len
32768
Questions
7250 (index 0–7249, full split)
Total rows
87000 (7250 × 12)
Generated by complete_prefix_vllm.py… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/code_rose_initial_1_7B_SFT_10K_rollouts_Qwen3-4B-Thinking-2507_k12_t0.7_maxtok12288.rosetta-ko-instruction-following-synth-sft-think
rosetta-ko-instruction-following-synth-sft-think
Korean-native instruction-following data — instructions with verifiable constraints (length, format, keywords, JSON, ...) checked by programmatic verifiers.
Synthetic data generated with the Qwen3.6-27B teacher model — part of the Rosetta-KO suite for the Rosetta Korean LLM (PoSTMEDIA).
Instruction-Following Suite
Sibling datasets from the same pipeline (each a separate repo):
repo
format… See the full description on the dataset page: https://huggingface.co/datasets/PoSTMEDIA/rosetta-ko-instruction-following-synth-sft-think.