VardanSharma13/spectrum-sft

Spectrum English SFT mix

Built by Vardan Sharma (MSc AI, University of Edinburgh) for the Spectrum / Spectrumize project with Vivek Iyer and Prof. Alexandra Birch.

English supervised fine-tuning data for a language model that reads its context through SONAR sentence embeddings and answers in text: reading comprehension (extractive, abstractive, boolean and multi-hop QA), multiple choice, NLI, classification, sentiment, toxicity, rating and summarisation, from 85 public datasets.

split	examples
train	1,295,303
validation	90,396
test	80,662

The train contexts hold about 27M sentences / 408M words.

Columns

column	meaning
context	the text the model reads (through SONAR)
task_type	EXTRACTIVE_QA, ABSTRACTIVE_QA, BOOLEAN_QA, MULTIHOP_QA, MCQ, NLI, CLASSIFICATION, SENTIMENT, TOXICITY, RATING, EDU_CLASSIFICATION, SUMMARIZATION
prompt_question	the question (empty for tasks without one, e.g. summarisation)
prompt_choices	answer options, without "A."-style labels; empty list for open tasks
target	the answer; always one of prompt_choices when choices exist (checked for every row)
dataset, split, uid	source dataset, split, unique id <dataset>:<source id>

No prompt templates are included: they are added at training time. Use dataset / task_type to reweight the mix.

from datasets import load_dataset
ds = load_dataset("VardanSharma13/spectrum-sft", split="train")

How it was built

1. Selection: 85 English datasets chosen from 104 candidates; per-dataset caps keep the mix balanced. Multi-hop QA (HotpotQA, 2WikiMultihopQA, MuSiQue) and MS MARCO contexts contain only the supporting / relevant passages, no distractors. Datasets without a test split get a 3% held-out test set.
2. Within each dataset: duplicates removed; train examples whose context appears in that dataset's validation/test removed.
3. SONAR-fragile lines (as in our pretraining data, following the LCM paper): URL lines, reference lines, long ID tokens, number/table rows and menu runs are removed from contexts. This is answer-safe: if a removed line holds the target or a choice, or if more than 30% of the context would go, the original context is kept. Contexts cleaned: 45,052 of 1,474,049 (kept unchanged to protect the answer: 163; too much removed: 3,182).
4. Across datasets: exact duplicates of (context, question, choices, target) removed (7,303 in train); train rows whose context also appears in any validation/test split removed (13).
5. Decontamination: train contexts sharing a 13-gram with Belebele, SIB-200, HellaSwag validation (the source of X-HellaSwag) or English XStoryCloze are removed (0 found after the earlier build-time pass).
6. Contract check: every row is validated (types, non-empty context and target, target in choices, no option labels, no "nan" strings, unique uid). SHA256SUMS.sft lists a checksum for every file.

Note on WikiLingua: wiki_lingua_en is part of the mix. If you evaluate auto-encoding on WikiLingua documents, check them against this train split or drop the dataset (dataset == "wiki_lingua_en").

Task types

task type	train	validation	test
SUMMARIZATION	218,452	17,219	21,609
EXTRACTIVE_QA	188,758	15,315	767
ABSTRACTIVE_QA	151,381	10,172	4,255
MCQ	140,378	15,608	7,886
CLASSIFICATION	132,878	4,286	14,659
NLI	127,990	10,539	13,829
MULTIHOP_QA	89,704	6,000	0
SENTIMENT	61,785	1,904	7,051
TOXICITY	60,681	5,275	5,434
RATING	53,712	0	3,142
BOOLEAN_QA	50,173	4,078	1,441
EDU_CLASSIFICATION	19,411	0	589

Datasets

dataset	train	validation	test
20_newsgroups	10,821	0	2,000
2wikimultihopqa	29,924	2,000	0
adversarial_qa	19,813	2,000	0
aeslc	13,498	1,825	1,771
ag_news	14,998	0	2,000
amazon_polarity	15,000	0	2,000
amazon_reviews_2023	58,174	0	1,824
anli	21,349	2,000	2,000
app_reviews	9,734	0	266
arxiv_class	2,266	286	298
arxiv_sum	7,322	1,466	1,490
bbc_news	1,147	0	974
big_patent	9,997	2,000	2,000
billsum	18,020	0	2,000
booksum	5,116	754	911
boolq	8,396	2,000	0
cb	250	56	0
civil_comments	14,998	2,000	2,000
cnn_dailymail	29,976	2,000	2,000
contract_nli	6,175	921	1,829
coqa	28,789	2,000	0
cosmos_qa	19,984	2,000	0
covid_qa	916	0	9
dbpedia_14	15,000	0	2,000
dialogsum	12,389	500	1,489
docnli	28,309	2,000	2,000
dream	6,113	2,000	2,000
drop	29,956	2,000	0
duorc	19,941	2,000	2,000
fineweb_edu_annot	19,411	0	589
govreport	2,114	56	34
hate_speech_offensive	19,384	0	616
hotpot_qa	39,852	2,000	0
imdb	14,997	0	2,000
ledgar	14,945	2,000	2,000
logiqa	7,316	640	624
mctest	1,096	200	600
mnli	19,988	2,000	2,000
mrqa	59,207	3,000	0
ms_marco	39,986	2,000	0
multi_news	14,993	2,000	2,000
multirc	19,543	2,000	0
musique	19,928	2,000	0
mutual	7,054	884	0
narrativeqa	32,347	2,000	2,000
news_category	14,561	0	439
openbookqa	2,080	448	443
patent_class	14,966	2,000	2,000
pubmed_qa	981	0	19
pubmed_qa_art	19,399	0	601
pubmed_sum	9,997	2,000	2,000
qasc	8,134	926	0
qasper	1,362	594	880
quac	30,000	2,000	0
quail	10,234	2,000	556
quality	809	741	0
quartz	2,696	384	784
quoref	19,384	2,000	0
race	39,853	2,000	2,000
reclor	4,591	500	0
record	19,997	2,000	0
reddit_tifu	14,582	0	418
ropes	10,924	1,688	0
rotten_tomatoes	7,709	1,040	1,051
rte	2,207	277	0
sciq	10,421	885	879
scitail	19,717	1,285	2,000
scitldr	1,988	618	618
snli	10,000	2,000	2,000
squad	29,759	2,000	0
squad_v2	13,069	1,630	0
sst2	9,101	864	0
strategyqa	1,602	0	687
subjqa_books	402	88	102
tweet_eval_offensive	11,301	1,275	818
tydiqa_en	3,685	440	0
wanli	19,995	0	2,000
wiki_lingua_en	28,515	0	878
wiki_qa	851	125	165
wiki_toxic	14,998	2,000	2,000
xlsum_en	19,979	2,000	2,000
xsum	29,966	2,000	2,000
yahoo_answers_topics	15,000	0	2,000
yelp_polarity	14,978	0	2,000
yelp_review_full	14,978	0	2,000

Licences

Every example keeps the licence of its source dataset (see the dataset's page on the Hugging Face Hub). Several sources are non-commercial (e.g. ANLI, ContractNLI, DialogSum, QuAIL, SciQ, WikiLingua, XL-Sum), so the mix as a whole is for research use. stats.json has the full build statistics.