README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: result
      dtype: string
    - name: source
      dtype: string
  splits:
    - name: act_en_news
      num_bytes: 994663
      num_examples: 998
    - name: no_index_en_news
      num_bytes: 1182747
      num_examples: 999
    - name: sample
      num_bytes: 15135
      num_examples: 27
    - name: malay_news
      num_bytes: 54838
      num_examples: 80
    - name: hard_en_news
      num_bytes: 2927696
      num_examples: 459
    - name: no_index_ms_news
      num_bytes: 1229858
      num_examples: 999
    - name: link_news
      num_bytes: 309948
      num_examples: 483
    - name: hard_ms_news
      num_bytes: 1716653
      num_examples: 211
    - name: akta_ms_news
      num_bytes: 1146385
      num_examples: 986
    - name: no_index_en_news_part2
      num_bytes: 2292171
      num_examples: 2325
    - name: no_index_ms_news_part2
      num_bytes: 3582293
      num_examples: 3659
    - name: news_emoji
      num_bytes: 2011913
      num_examples: 3384
    - name: malaysia_hansard
      num_bytes: 3586533
      num_examples: 4989
    - name: chat_assistant
      num_bytes: 13104843
      num_examples: 18646
    - name: chat_assistant_simple
      num_bytes: 5147784
      num_examples: 7507
    - name: pseudolabel_en_Qwen3_8B
      num_bytes: 33839354
      num_examples: 48055
    - name: pseudolabel_ms_Qwen3_8B
      num_bytes: 75517371
      num_examples: 117506
    - name: pseudolabel_ms_akta_Qwen3_8B
      num_bytes: 22537967
      num_examples: 29106
    - name: pseudolabel_en_part2_Qwen3_8B
      num_bytes: 36895082
      num_examples: 50360
    - name: pseudolabel_chat_assistant_Qwen3_8B
      num_bytes: 480703638
      num_examples: 658925
    - name: pseudolabel_malay_news_Qwen3_8B
      num_bytes: 750721011
      num_examples: 1073163
    - name: pseudolabel_english_news_Qwen3_8B
      num_bytes: 644877397
      num_examples: 904559
  download_size: 868334534
  dataset_size: 2102447862
configs:
  - config_name: default
    data_files:
      - split: act_en_news
        path: data/act_en_news-*
      - split: no_index_en_news
        path: data/no_index_en_news-*
      - split: sample
        path: data/sample-*
      - split: malay_news
        path: data/malay_news-*
      - split: hard_en_news
        path: data/hard_en_news-*
      - split: akta_ms_news
        path: data/akta_ms_news-*
      - split: no_index_ms_news
        path: data/no_index_ms_news-*
      - split: link_news
        path: data/link_news-*
      - split: hard_ms_news
        path: data/hard_ms_news-*
      - split: malaysia_hansard
        path: data/malaysia_hansard-*
      - split: no_index_en_news_part2
        path: data/no_index_en_news_part2-*
      - split: no_index_ms_news_part2
        path: data/no_index_ms_news_part2-*
      - split: news_emoji
        path: data/news_emoji-*
      - split: chat_assistant
        path: data/chat_assistant-*
      - split: chat_assistant_simple
        path: data/chat_assistant_simple-*
      - split: pseudolabel_en_Qwen3_8B
        path: data/pseudolabel_en_Qwen3_8B-*
      - split: pseudolabel_ms_Qwen3_8B
        path: data/pseudolabel_ms_Qwen3_8B-*
      - split: pseudolabel_ms_akta_Qwen3_8B
        path: data/pseudolabel_ms_akta_Qwen3_8B-*
      - split: pseudolabel_en_part2_Qwen3_8B
        path: data/pseudolabel_en_part2_Qwen3_8B-*
      - split: pseudolabel_chat_assistant_Qwen3_8B
        path: data/pseudolabel_chat_assistant_Qwen3_8B-*
      - split: pseudolabel_malay_news_Qwen3_8B
        path: data/pseudolabel_malay_news_Qwen3_8B-*
      - split: pseudolabel_english_news_Qwen3_8B
        path: data/pseudolabel_english_news_Qwen3_8B-*

Malaysian Normalizer

Normalize numbers, digits, currency, IC, time, date, timestamp, email, URL, titles, abbrevations and symbols.

Instruction format

1. We make sure the normalized text able to reverse back to the original text using the normalized mapping.
2. We make sure the normalized text does not contain any digits.
3. We predict major language in the normalized mapping and use it as prompt language.
4. Convert to instruction format, uploaded at mesolitica/Malaysian-SFT/malaysian_normalizer,

Pseudolabel

1. We finetuned initial version on Qwen/Qwen3-8B, the model released at malaysia-ai/Malaysian-Normalizer-Qwen3-8B.
2. We massively pseudolabel and use the same verifier steps from instruction format.
3. Convert to instruction format, uploaded at mesolitica/Malaysian-SFT/malaysian_normalizer_pseudolabel.