Datasets:

SultanR
/

openthoughts3-en-ar-midtrain

row_id stringlengths 28 28	messages listlengths 2 2	ar_messages listlengths 2 2	src_dataset stringclasses 1 value	src_config stringclasses 1 value	src_split stringclasses 1 value	src_file stringclasses 4 values	src_row_index int64 0 10k	source stringclasses 1 value	n_hard_breaks_restored int64 0 512	n_chunks int64 6 133	n_fallback_chunks int64 0 43	think_present bool 1 class	think_closed bool 2 classes	qc_arabic_fraction float64 0.05 1	qc_source_repeated_ngram int64 1 11.7k	qc_numbers_match bool 2 classes	qc_max_repeated_ngram int64 1 9.18k	qc_clean bool 2 classes	pipeline_version stringclasses 1 value	prompt_version stringclasses 1 value	model stringclasses 1 value	model_revision stringclasses 1 value
openthoughts3:00000:00000000	[ { "role": "user", "content": "I am in desperate need of some ear defenders, so I can program in peace. Unfortunately, I don't have ear defenders. But I do have a pair of headphones, a microphone and a microcontroller, so I thought I'd make some noise-cancelling headphones.\nHowever, there is one tiny proble...	[ { "role": "user", "content": "أنا في حاجة ماسة إلى بعض واقيات الأذن، حتى أتمكن من البرمجة بسلام. لسوء الحظ، لا أملك واقيات أذن. لكن لدي زوج من سماعات الرأس، وميكروفون، ومتحكم دقيق، لذا فكرت في صنع سماعات رأس عازلة للضوضاء.\nومع ذلك، هناك مشكلة صغيرة واحدة. لا يمكنني البرمجة في البيئات الصاخبة! لذا ستحتاج إل...	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	0	OpenThoughts3-1.2M	0	66	2	true	true	0.7501	5	true	5	false	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000001	[{"role":"user","content":"You are to golf a program that will take a filename as input and you must(...TRUNCATED)	[{"role":"user","content":"يجب عليك كتابة برنامج بأقل عدد ممكن من ا(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	1	OpenThoughts3-1.2M	0	81	0	true	true	0.7768	7	true	7	true	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000002	[{"role":"user","content":"Given a set of substrings, such as [ca, ar, car, rd], it's possible to cr(...TRUNCATED)	[{"role":"user","content":"بفرض وجود مجموعة من السلاسل الفرعية، مث(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	2	OpenThoughts3-1.2M	0	81	2	true	true	0.865	3	true	3	false	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000003	[{"role":"user","content":"Jury picked a polynomial f(x) = a_0 + a_1 ⋅ x + a_2 ⋅ x^2 + ... + a_k(...TRUNCATED)	[{"role":"user","content":"اختارت لجنة متعددة الحدود f(x) = a_0 + a_1 ⋅ x + (...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	3	OpenThoughts3-1.2M	0	80	0	true	false	0.8773	3	true	2	true	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000004	[{"role":"user","content":"The Dutch system for naming one's ancestors is a lot more interesting tha(...TRUNCATED)	[{"role":"user","content":"إن النظام الهولندي لتسمية أسلاف المرء أ(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	4	OpenThoughts3-1.2M	0	79	1	true	false	0.7932	2	true	2	false	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000005	[{"role":"user","content":"\nInspired by this\n\nYour task today: given two strings, find the string(...TRUNCATED)	[{"role":"user","content":"\nمستوحى من هذا\n\nمهمتك اليوم: بمعطى سلسل(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	5	OpenThoughts3-1.2M	0	80	0	true	true	0.8769	2	true	2	true	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000006	[{"role":"user","content":"Let's define a function \\$f\\$ which, given a positive integer \\$x\\$, (...TRUNCATED)	[{"role":"user","content":"لنُعرّف دالة $f\\$ والتي، بمعطى عدد صحيح م(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	6	OpenThoughts3-1.2M	4	53	0	true	true	0.8367	4	true	3	true	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000007	[{"role":"user","content":"You are given a positive integer k and an array a_1, a_2, …, a_n of non(...TRUNCATED)	[{"role":"user","content":"يُعطى لك عدد صحيح موجب k ومصفوفة $a_1, a_2, \\do(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	7	OpenThoughts3-1.2M	0	81	0	true	false	0.8738	2	true	2	true	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000008	[{"role":"user","content":"F: Red-Black Soul Gem\n\nproblem\n\nHomu-chan has the mysterious power to(...TRUNCATED)	[{"role":"user","content":"F: Red-Black Soul Gem\n\nالمسألة\n\nتمتلك Homu-chan القو(...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	8	OpenThoughts3-1.2M	0	81	0	true	false	0.8608	6	true	5	true	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
openthoughts3:00000:00000009	[{"role":"user","content":"A permutation of size n is an array of size n such that each integer from(...TRUNCATED)	[{"role":"user","content":"التبديلة (permutation) من الحجم n هي مصفوفة من (...TRUNCATED)	HuggingFaceTB/smoltalk2	Mid	openthoughts3	Mid/OpenThoughts3_1.2M-00000-of-00113.parquet	9	OpenThoughts3-1.2M	0	77	2	true	true	0.7358	3	true	3	false	1.0.0	3efd6506ccf0	google/gemma-4-12B-it	707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7

End of preview. Expand in Data Studio

openthoughts3-en-ar-midtrain

Arabic translation of the OpenThoughts3_1.2M split of smoltalk2 (config Mid): long mathematical reasoning traces with <think> blocks, in a two-message user/assistant format. Translated with google/gemma-4-12B-it (bf16, greedy) on A100s. All 1,135,104 source rows are present, none dropped.

The pipeline segments each message into prose and verbatim blocks (code, LaTeX, tables, and inline non-translatables are masked and never sent to the model), translates prose in ~256-token chunks, and retries failed chunks once before falling back to their English source rather than dropping them. Nothing is silently repaired, every row carries QC labels. A sibling corpus from the same pipeline is available at nemotron-mc-en-ar-midtrain.

Details


Rows	1,135,104 (equals the source split exactly)
Source tokens	16.74B
Chunks translated	81.57M (median 78 per row)
Size	41.5 GB
`qc_clean`	93.12%
Chunk-level fallback rate	0.15%
AR/EN character ratio	0.923 mean
`<think>` tag count EN vs AR mismatches	0 rows

62% of rows have an unclosed <think> block. This is a property of the upstream English source, present there too. Filter on think_closed if you need complete traces.

When comparing quality to other corpora, use the chunk-level fallback rate, not row-level rates. At ~72 chunks per row, one bad chunk flags the whole row, so a 0.15% chunk rate arithmetically becomes a ~6% rows-with-fallback rate.

Schema

Column	Description
`row_id`	Position in the pinned source snapshot, maps back to the exact source row
`messages`	English source conversation, unmodified
`ar_messages`	Arabic conversation
`source`	Upstream `source` field, passed through
`src_dataset`, `src_config`, `src_split`, `src_file`, `src_row_index`	Provenance
`n_chunks`, `n_fallback_chunks`, `n_masks_lost`, `n_masks_duplicated`, `n_hard_breaks_restored`	Translation counters
`think_present`, `think_closed`	Reasoning trace structure
`qc_clean`, `qc_arabic_fraction`, `qc_numbers_match`, `qc_numbers_missing`, `qc_max_repeated_ngram`, `qc_think_open_delta`, `qc_think_close_delta`	Quality labels
`pipeline_version`, `prompt_version`, `model`, `model_revision`	Reproducibility

qc_clean is the conjunction of the structural checks: no placeholder lost, no English fallback, <think> structure intact, output is Arabic, no decode loop. It is a structural claim, not a fluency judgement. A chunk that failed translation twice keeps its English source, so some rows contain English spans by design; sampled instances are overwhelmingly code, which is correct behavior. The Arabic character fraction is below 1.0 by design, since identifiers, units, and LaTeX are preserved in Latin script.

Usage

from datasets import load_dataset

ds = load_dataset("SultanR/openthoughts3-en-ar-midtrain", split="train", streaming=True)
clean = ds.filter(lambda r: r["qc_clean"] and r["think_closed"])

License

Refer to the smoltalk2 source dataset for license information.

Downloads last month: 78

Collection including SultanR/openthoughts3-en-ar-midtrain

AraTrain

Collection

Collection of Arabic training datasets • 12 items • Updated 3 days ago