mlg-post-training / README.md
SaarAI's picture
Malagasy post-training data (Walia format)
479c1db verified
|
Raw History Blame Contribute Delete
10.5 kB
metadata
language:
  - mg
  - plt
  - en
task_categories:
  - text-generation
  - translation
  - text-classification
  - token-classification
  - summarization
pretty_name: Malagasy post-training (Walia-style instruction data)
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

mlg-post-training

Malagasy instruction / chat data compiled from public datasets in the style of Walia-LLM (Azime et al., 2024, arXiv:2402.08015), in the same row format as SaarAI/zul-post-training, SaarAI/kin-post-training and SaarAI/amh-post-training: task datasets (translation, classification, tagging, NER, generation) are converted to instructions using paraphrased templates (templates.json), and existing instruction / chat sets are added as-is.

576,271 train / 2,932 validation rows. Validation is a random per-task holdout (0.5%, 20-1000 rows per task).

Columns

id, source, task, instruction_lang (mlg or en = language of the template, native = source already an instruction/chat set), instruction, input, output, messages (chat form: list of role/content dicts, multi-turn for chat sources; for multi-turn rows instruction / output are the first user / assistant turn).

Tasks

task rows
translation_mlg_en 112,034
translation_en_mlg 112,034
question_answering 57,580
instruction_following 54,475
text_continuation 39,995
chat 37,483
generation 24,510
title_generation 21,368
summarization 19,933
poem_generation 16,269
code_chat 12,200
dialogue 9,999
text_simplification 9,991
event_linking 9,949
paraphrase_identification 9,918
aya_human 9,649
article_generation 6,368
math_reasoning 5,889
sentiment_classification 4,536
pos_tagging 4,227
topic_classification 701
named_entity_recognition 95

Sources

source key dataset note
smol_sent, smol_doc google/smol (smolsent / smoldoc en_mg) professional en->mg sentence and document translations
opus_globalvoices wecover/OPUS_GlobalVoices (en-mg train) Global Voices news sentence pairs
polynews_gv aiana94/polynews-parallel (eng_Latn-plt_Latn) Global Voices news pairs (100k sample)
sib200 Davlan/sib200 (plt_Latn train) topic classification (7 classes)
vaovao_sentiment Lo-Renz-O/vaovao_malagasy_sentiment_corpus (train) news-sentence sentiment, native-annotated (2 classes)
sokajy_pos Vatosoa/pos-tagging-malagasy-sokajy expert-annotated UPOS tagging (CoNLL-U)
wikiann_mg unimelb-nlp/wikiann (mg train) named entity recognition (PER / LOC / ORG)
wikipedia_mg wikimedia/wikipedia (20231101.mg) Wikipedia -> article / title generation (>= 100 words)
news_mg malagasy-asr/malagasy-text-corpus (tia_tanindrazana.json, news_mada.json, gvalosoa.json) news -> headline generation (15k sample)
tononkalo Lo-Renz-O/Collection-Tononkalo-Malagasy poems (Vetso Serasera) -> poem generation
fineweb2_mlg SaarAI/fineweb-2-malagasy (plt_Latn train) web text -> text continuation (40k sample)
wikiconversations_mlg israel/WikiConversations (malagasy) multi-turn conversations grounded in Wikipedia
kakugo_plt ptrdvn/kakugo-plt synthetic chat; <think> blocks removed
code170k_mlg michsethowusu/Code-170k-malagasy machine-translated code Q&A chat
alpaca_mlg saillab/alpaca-malagasy-cleaned machine-translated Alpaca
dolly_mlg saillab/taco-datasets (Dolly-15k Malagasy) machine-translated Dolly
gsm8k_mlg Lo-Renz-O/grade-school-math-instructions-Malagasy translated GSM8K; GSM8K-test problems removed
aya_* CohereLabs/aya_collection_language_split (plateau_malagasy train) Aya mix (translated); each sub-dataset capped at 10k

Row counts per source key:

source rows
polynews_gv 184,128
alpaca_mlg 40,694
fineweb2_mlg 39,995
opus_globalvoices 37,272
kakugo_plt 37,092
tononkalo 16,269
news_mg 15,000
dolly_mlg 13,781
wikipedia_mg 12,736
code170k_mlg 12,200
aya_SODA_inst_T 9,999
aya_Wiki_split_inst_T 9,991
aya_PIQA_T 9,972
aya_CNN_Daily_Mail_T 9,972
aya_NQ_Open_T 9,971
aya_Flan_GEM_wiki_lingua_T 9,961
aya_Flan_CoT_submix_T 9,958
aya_Xlel_wd_inst_T 9,949
aya_Adversarial_QA_T 9,938
aya_PAWS_Wiki_T 9,918
aya_Mintaka_inst_T 9,904
aya_HotpotQA_T 9,814
aya_Aya_Dataset 9,649
aya_Dolly_v2_T 9,548
aya_Flan_Coqa_T 6,406
gsm8k_mlg 5,889
vaovao_sentiment 4,536
aya_Flan_lambada_T 4,269
sokajy_pos 4,227
smol_sent 1,720
aya_WIKI_QA_T 1,037
smol_doc 948
aya_Joke_explaination_inst_T 735
sib200 701
aya_Flan_unified_QA_T 538
wikiconversations_mlg 391
wikiann_mg 95

Filtering

  • Chat / instruction sources: system turns dropped, <think> blocks removed, conversations with an empty or literal "None" turn dropped, decoding-collapse (a word repeated 8+ times in a row, or very low compressibility) dropped, and GlotLID (cis-lmu/glotlid) must give Malagasy (plt_Latn + skg_Latn + xmv_Latn + tdx_Latn) >= 0.5 on the assistant text (code chat excluded).
  • Translation pairs: mg/en character-length ratio 0.5-3.0, at least 2 words per side, identical pairs dropped, and the Malagasy side must pass the same GlotLID check.
  • gsm8k_mlg covers both GSM8K train and test; every problem whose number sequence matches a GSM8K test question is removed, so MGSM / AfriMGSM stay clean.
  • 10,618 exact duplicates were removed (same task + input/instruction + output; full conversation for chat).
reason rows removed
aya_SODA_inst_T: over cap 10000 1,181,582
aya_Wiki_split_inst_T: over cap 10000 979,944
aya_Xlel_wd_inst_T: over cap 10000 513,112
aya_HotpotQA_T: over cap 10000 345,476
polynews_gv: over cap 100000 223,288
aya_NQ_Open_T: over cap 10000 165,850
aya_CNN_Daily_Mail_T: over cap 10000 90,000
aya_Flan_CoT_submix_T: over cap 10000 81,910
wikipedia_mg: < 100 words 64,883
aya_PAWS_Wiki_T: over cap 10000 39,401
news_mg: no headline / body < 80 words 29,652
wikipedia_mg: bot-stub template (> 10 articles share its opening) 25,065
aya_Flan_GEM_wiki_lingua_T: over cap 10000 17,147
aya_PIQA_T: over cap 10000 6,113
aya_Dolly_v2_T: over cap 10000 4,808
aya_Aya_Dataset: over cap 10000 4,597
news_mg: over cap 15000 4,268
aya_Mintaka_inst_T: over cap 10000 4,000
kakugo_plt: not Malagasy (GlotLID) 3,713
gsm8k_mlg: matches a GSM8K test problem (decontamination) 2,223
polynews_gv: Malagasy side not Malagasy (GlotLID) 2,048
tononkalo: no title / < 30 or > 800 words 1,298
dolly_mlg: not Malagasy (GlotLID) 1,207
polynews_gv: bad pair (length ratio / too short / identical) 957
alpaca_mlg: not Malagasy (GlotLID) 881
opus_globalvoices: Malagasy side not Malagasy (GlotLID) 742
opus_globalvoices: bad pair (length ratio / too short / identical) 612
gsm8k_mlg: not Malagasy (GlotLID) 584
aya_Dolly_v2_T: not Malagasy (GlotLID) 415
sokajy_pos: < 2 tokens 355
aya_HotpotQA_T: not Malagasy (GlotLID) 121
aya_Aya_Dataset: not Malagasy (GlotLID) 88
aya_Adversarial_QA_T: decoding collapse 40
aya_Flan_CoT_submix_T: not Malagasy (GlotLID) 39
aya_Xlel_wd_inst_T: not Malagasy (GlotLID) 33
code170k_mlg: decoding collapse 29
aya_Dolly_v2_T: decoding collapse 28
aya_Flan_GEM_wiki_lingua_T: decoding collapse 28
aya_PIQA_T: decoding collapse 18
alpaca_mlg: decoding collapse 16
aya_NQ_Open_T: not Malagasy (GlotLID) 16
aya_Flan_GEM_wiki_lingua_T: not Malagasy (GlotLID) 11
aya_Flan_lambada_T: not Malagasy (GlotLID) 10
aya_PAWS_Wiki_T: not Malagasy (GlotLID) 10
aya_CNN_Daily_Mail_T: decoding collapse 9
aya_Xlel_wd_inst_T: decoding collapse 9
kakugo_plt: malformed / empty / None turn 6
aya_PIQA_T: not Malagasy (GlotLID) 6
aya_Wiki_split_inst_T: decoding collapse 6
aya_Xlel_wd_inst_T: malformed / empty / None turn 6
aya_Adversarial_QA_T: not Malagasy (GlotLID) 5
aya_Mintaka_inst_T: not Malagasy (GlotLID) 5
kakugo_plt: decoding collapse 4
dolly_mlg: decoding collapse 4
code170k_mlg: malformed / empty / None turn 3
aya_Flan_CoT_submix_T: decoding collapse 3
aya_Flan_Coqa_T: not Malagasy (GlotLID) 3
aya_Wiki_split_inst_T: not Malagasy (GlotLID) 3
smol_sent: Malagasy side not Malagasy (GlotLID) 2
aya_Flan_unified_QA_T: not Malagasy (GlotLID) 2
aya_WIKI_QA_T: not Malagasy (GlotLID) 2
smol_sent: bad pair (length ratio / too short / identical) 1
wikiconversations_mlg: not Malagasy (GlotLID) 1
aya_Aya_Dataset: malformed / empty / None turn 1
aya_CNN_Daily_Mail_T: not Malagasy (GlotLID) 1
aya_SODA_inst_T: decoding collapse 1

Known limitations (please read)

  • The Malagasy instruction templates (templates.json) and translated label names were drafted by an AI model and have not been reviewed by a native speaker.
  • Quality varies by source: code170k_mlg, alpaca_mlg, dolly_mlg, gsm8k_mlg and the aya_* (T) sets are machine-translated; kakugo_plt is synthetic. Use the source column to filter or reweight.
  • sokajy_pos has no official split and its first sentences are Bible verses; WikiANN mg train has only 100 sentences.
  • opus_globalvoices and polynews_gv are both Global Voices and may overlap (exact duplicates are removed).
  • Excluded after inspection: Helsinki-NLP/opus-100 en-mg (software UI strings), michsethowusu/english-malagasy_sentence-pairs and michsethowusu/malagasy-sentiments-corpus / -emotions-corpus (mined pairs / auto-labels, many misaligned), Maminirina1/MalagasyEnglish (noisy mined pairs, many Bible / subtitles), the mt560 sets (tokenized, mostly JW / Bible), FLORES / Belebele (benchmarks). The gated Lo-Renz-O/alpaca-gpt4-Malagasy was not accessible.
  • Only train splits were used; test splits (SIB-200, VMSC validation, WikiANN, FineWeb-2 plt_Latn test) remain usable for evaluation. Licenses differ per source (Alpaca / Dolly / Aya derived data carry upstream terms); check before commercial use.