Omarrran/Global_Setu

Global Setu

One English sentence, 57 aligned translations (22 Indic + 35 global), usable as any-to-any pairs.

Note. Machine-translated research corpus, not human-verified gold.

Overview

984,764 English rows by 60 columns: id, english, source, 22 Sarvam Indic translations and 35 Riva global translations. All translations in a row align to that row's English input.

Rows	984,764 (IDs 0-984763)
Columns	3 identity + 22 Sarvam_* + 35 Riva_*
Target columns	57, covering 56 languages
Hindi	Kept in both families: Sarvam_Hindi and Riva_Hindi
Croatian	Intentionally excluded (not a null column)
Viewer / load_dataset files	data/train-*.parquet (62 Zstandard shards)

Any-to-any translation

Every target column is translated from the same English sentence, so each row is a multiway-aligned record. Any two columns form a translation pair, for example Marathi to German (Sarvam_Marathi to Riva_German), without needing direct Marathi-German data. English is the pivot.

Aligned text columns	58 (English + 57 targets)
Directed pairs, English included	3,306 (58 x 57)
Directed pairs, targets only	3,192 (57 x 56)
Rows per pair	up to 984,763 (row 934791 excluded)

• Non-English pairs are pivot-derived and synthetic. Errors from both machine translations can compound, so a pair is only as reliable as its weaker side.
• Sarvam_Hindi and Riva_Hindi are two machine translations of the same sentence, so that pair is a paraphrase pair, not a cross-language pair.
• This dataset has no human-corrected columns today. If corrected columns are added, they would carry through to every pair that uses them, giving a more reliable route between any two languages.

Exceptional row. ID 934791 has english = null. Its 22 Sarvam cells are kept, but all 35 Riva cells are null. Filter it out when you need non-null English.

Schema

Column family	Count	Non-null rows per column
id, english, source	3	English: 984,763
Sarvam_<Language>	22	984,764
Riva_<Language>	35	984,763

Parallel any-to-any dataset: languages available (57).

English	Assamese	Bengali	Bodo	Dogri
Gujarati	Hindi	Kannada	Kashmiri	Konkani
Maithili	Malayalam	Manipuri	Marathi	Nepali
Odia	Punjabi	Sanskrit	Santali	Sindhi
Tamil	Telugu	Urdu	Czech	Danish
German	Greek	Spanish (Spain)	Spanish (Latin America)	Finnish
French	Hungarian	Italian	Lithuanian	Latvian
Dutch	Norwegian	Polish	Portuguese (Portugal)	Portuguese (Brazil)
Romanian	Russian	Slovak	Swedish	Chinese (Simplified)
Chinese (Traditional)	Japanese	Korean	Estonian	Slovenian
Bulgarian	Ukrainian	Arabic	Vietnamese	Turkish
Indonesian	Thai

Sarvam columns (22)

Sarvam_Assamese, Sarvam_Bengali, Sarvam_Bodo, Sarvam_Dogri, Sarvam_Gujarati, Sarvam_Hindi, Sarvam_Kannada, Sarvam_Kashmiri, Sarvam_Konkani, Sarvam_Maithili, Sarvam_Malayalam, Sarvam_Manipuri, Sarvam_Marathi, Sarvam_Nepali, Sarvam_Odia, Sarvam_Punjabi, Sarvam_Sanskrit, Sarvam_Santali, Sarvam_Sindhi, Sarvam_Tamil, Sarvam_Telugu, Sarvam_Urdu

Riva columns (35)

Riva_Czech, Riva_Danish, Riva_German, Riva_Greek, Riva_Spanish (Spain), Riva_Spanish (Latin America), Riva_Finnish, Riva_French, Riva_Hungarian, Riva_Italian, Riva_Lithuanian, Riva_Latvian, Riva_Dutch, Riva_Norwegian, Riva_Polish, Riva_Portuguese (Portugal), Riva_Portuguese (Brazil), Riva_Romanian, Riva_Russian, Riva_Slovak, Riva_Swedish, Riva_Chinese (Simplified), Riva_Chinese (Traditional), Riva_Japanese, Riva_Hindi, Riva_Korean, Riva_Estonian, Riva_Slovenian, Riva_Bulgarian, Riva_Ukrainian, Riva_Arabic, Riva_Vietnamese, Riva_Turkish, Riva_Indonesian, Riva_Thai

Some Riva names contain spaces and parentheses. Quote them in SQL ("Riva_Spanish (Spain)") and use bracket access in Python.

Loading

Log in first: hf auth login.

from datasets import load_dataset

repo = "Omarrran/Global_Setu"
ds = load_dataset(repo, split="train", token=True)
print(ds[0]["english"], ds[0]["Sarvam_Kashmiri"], ds[0]["Riva_German"])

Stream, or load only the columns you need:

stream = load_dataset(repo, split="train", streaming=True, token=True)
for row in stream.take(3):
    print(row["id"], row["english"], row["Riva_French"])

ds = load_dataset(repo, split="train", token=True,
                  columns=["id", "english", "Sarvam_Kashmiri", "Riva_German"])

Select targets, compare models, or take a whole family:

compare_hindi = ds.select_columns(["id", "english", "Sarvam_Hindi", "Riva_Hindi"])

sarvam_cols = [c for c in ds.column_names if c.startswith("Sarvam_")]
riva_cols = [c for c in ds.column_names if c.startswith("Riva_")]

Drop the null-English row:

clean = ds.filter(lambda r: r["english"] is not None)   # 984,763 rows

Pandas, Polars and DuckDB read the Parquet shards directly:

import pandas as pd
df = pd.read_parquet(
    "hf://datasets/Omarrran/Global_Setu/data/train-00000-of-00062.parquet",
    columns=["id", "english", "Riva_French"],
)

import polars as pl
out = (pl.scan_parquet("hf://datasets/Omarrran/Global_Setu/data/train-*.parquet")
         .select(["id", "english", "Sarvam_Telugu"])
         .filter(pl.col("english").is_not_null())
         .head(5).collect())

import duckdb
duckdb.sql("""
    SELECT id, english, "Riva_Chinese (Simplified)"
    FROM 'hf://datasets/Omarrran/Global_Setu/data/train-*.parquet'
    WHERE english IS NOT NULL LIMIT 5
""").show()

Download only what you need:

hf download Omarrran/Global_Setu --repo-type dataset \
  --include "data/*.parquet" --include "metadata.json" --local-dir ./global-setu

English-to-target records:

def pairs(rows, columns):
    for row in rows:
        if row["english"] is None:
            continue
        for col in columns:
            if row[col] is not None:
                yield {"id": row["id"], "english": row["english"],
                       "target_column": col, "translation": row[col]}

Any pair, for example Marathi to German:

def make_pair(repo, src, tgt):
    d = load_dataset(repo, split="train", token=True, columns=["id", src, tgt])
    d = d.filter(lambda r: r[src] is not None and r[tgt] is not None)
    return d.rename_columns({src: "source_text", tgt: "target_text"})

mr_de = make_pair(repo, "Sarvam_Marathi", "Riva_German")

All directions among a chosen set of columns:

from itertools import permutations

cols = ["english", "Sarvam_Marathi", "Riva_German", "Riva_Japanese"]
for src, tgt in permutations(cols, 2):
    pair = make_pair(repo, src, tgt)

Use Parquet for normal analysis. The JSONL (10.7 GB) and XLSX (4.9 GB) archives are for preservation only.

Files

Path	Details
data/train-*.parquet	62 Zstandard shards, 3,529,865,033 bytes total, largest 57,964,539 bytes
jsonl/GlobalSetu_22Sarvam_35Riva_1M.jsonl	Wide UTF-8 archive, 10,733,116,611 bytes, one object per row in ID order
full/GlobalSetu_22Sarvam_35Riva_1M.xlsx	Wide workbook, 4,913,563,752 bytes, 984,764 data rows plus header
metadata.json	Counts, schema, sizes, SHA-256 checksums
verification.json	Independent cross-format and workbook readback
excel_cell_adjustments.json	Affected IDs and columns for the Excel caveat

Validation

• Ordered source digest matched the published BharatSetu English, ID and provenance sequence.
• All 35 Riva shard sets complete with contiguous IDs.
• Verifier checked 984,764 JSONL rows, all 62 Parquet shards (schema, row counts, size caps) and 984,765 XLSX rows including the header.

These are structural checks, not a translation-accuracy score. No human-quality audit is claimed.

Excel caveat. Excel cannot store certain control characters, so they were removed from 14 XLSX cells (9 English, 5 Riva); no cells were truncated. JSONL and Parquet keep the exact original strings.

Limitations

• Machine translations may contain errors, hallucinations, inappropriate content or model bias.
• Non-English pairs are pivoted through English and have not been validated as direct translations; evaluate on human-verified data before relying on them.
• Source provenance may include material with differing third-party rights; review source and model terms before redistribution or commercial use.
• No license is asserted on behalf of upstream sources.

Citation

@dataset{global_setu,
  title     = {Global Setu: English-to-57 Machine Translation Columns},
  author    = {Haq Nawaz Malik},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Omarrran/Global_Setu}
}

Maintainer

Omar (Haq Nawaz Malik) · Hugging Face · GitHub