docs/DATASET_CARD_draft.md

metadata

language:
  - hi
  - en
  - multilingual
language_bcp47:
  - hi-Latn
license: apache-2.0
task_categories:
  - text-generation
tags:
  - hinglish
  - hindi
  - pretraining
  - code
  - science
  - legal
pretty_name: Viu Mini Raw Pretrain
size_categories:
  - 100M<n<1B
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - hindi/train-*.parquet
          - hindi_fixed/train-*.parquet
          - english_fixed/train-*.parquet
          - hinglish/train-*.parquet
          - domains/train-*.parquet
          - translation/train-*.parquet
          - translation/samanantar-*.parquet
          - wikipedia/train-*.parquet
          - science/*.parquet
          - math/*.parquet
          - code/*.parquet
          - legal/train-*.parquet
          - health/*.parquet
          - stories/*.parquet
          - grammar/*.parquet
          - finance/*.parquet
          - dictionary/*.parquet
          - toxicity/*.parquet
          - uncensored/*.parquet

ViuAI/viu-mini-raw-pretrain

Raw pretraining corpus for ViuMini-MoE-242M (Hinglish-first; language-mix rule OPTIONAL, default OFF — training config me enforce_mix dekho). Unified schema: text, lang, source, domain, safety_tag (older files still migrating — see Known issues).

• text: training text (plain, or <|user|> ... <|assistant|> ... templated for QA/instruct sources)
• lang: hinglish | hindi | english | bilingual
• source: origin dataset/generator name (e.g. indiccorp_v2, fineweb-edu, vyakaran_master, sciq)
• domain: knowledge pillar (e.g. general_hindi, electronics_smt_pcb, science_qa, hindi_vyakaran_sandhi)
• safety_tag: safe | toxic | uncensored

Safety notice (intentional inclusion)

toxicity/ and uncensored/ folders are intentionally included for robustness research (hate-speech detection, refusal training, red-teaming). They are TAGGED, not hidden:

from datasets import load_dataset
ds = load_dataset("ViuAI/viu-mini-raw-pretrain", split="train", streaming=True)

def keep(row, allow_toxic=False, allow_uncensored=False):
    tag = (row.get("safety_tag") or "safe").lower()
    if tag == "safe": return True
    if tag == "toxic": return allow_toxic
    if tag == "uncensored": return allow_uncensored
    return True

# default pretraining: safe only
ds_safe = ds.filter(lambda r: keep(r))
# robustness research: safe+toxic
ds_research = ds.filter(lambda r: keep(r, allow_toxic=True))
# red-team only, controlled: everything
ds_full = ds.filter(lambda r: keep(r, allow_toxic=True, allow_uncensored=True))

Do NOT train user-facing models on full without alignment (SFT/DPO + refusal).

Known issues (2026-09-22 audit, fixing in progress)

1. Hindi shards train-00115..00178, train-00307..00370 missing (upload gaps) — backfilling.
2. english/ 37x1.6GB + hindi/ 13x1.15GB oversized shards — resharding to ~150MB/100k rows.
3. Legacy schema drift (question/answer, instruction/response, topic/difficulty, title/text) — migrating to unified schema via data/scripts/standardize_to_unified.py.
4. Viewer HTTP 500 on some offsets until (2)+(3) complete.
5. Duplicate curated pillar rows (domains/ vs generated) — dedup via stable blake2b hash.

See docs/HF_REPAIR_RUNBOOK.md in the model repo for repair commands.

License / citation

Apache-2.0. Upstream sources retain their own licenses (Wikipedia CC-BY-SA, Samanantar CC-BY, etc.). If you use this dataset, cite upstream sources plus ViuAI/viu-mini-raw-pretrain.