💬 FineTranslations

The world's knowledge in 1+1T tokens of parallel text

What is it?

This dataset contains over 1 trillion tokens of parallel text in English and 500+ languages. It was obtained by translating data from 🥂 FineWeb2 into English using Gemma3 27B.

We relied on datatrove's inference runner to deploy a synthetic data pipeline at scale. Its checkpointing and VLLM lifecycle management features allowed us to use leftover compute from the HF cluster without fear of preemption. The async implementation ensures strong GPU utilization at all times.

The 💬 FineTranslations dataset is fully reproducible and available under the permissive ODC-By 1.0 license.

This is the base version. For the Edu version, see here.

What is it for?

The main motivation behind the creation of this dataset was improving translation capabilities. While models are generally strong at translating from other languages into English (X->English), the opposite is often not true, particularly for lower resource languages. Our approach was to take data that was originally in non-English languages (from 🥂 FineWeb2, our large multilingual pre-training dataset), chunk it, and translate it. This dataset can then be used to improve English->X translations by finetuning existing models (we leave this for future work).

Additionally, the resulting English data contains relevant cultural information for different countries and languages, and our experiments show that the 1T tokens we obtained perform on a similar level as our 🍷 FineWeb dataset. This data can therefore also be used for English only model training (potentially as an extension of FineWeb).

For the English only performance of the dataset, see the comparison below. The ablation setup is the same as in FinePDFs.

Languages and available subsets

Each language is identified by its ISO 639-3 code, and the data is grouped by language-script pairs, since some languages have content in multiple scripts.

The full list of subsets is available here.

To access data from all the languages, use the all subset.

How to download and use 💬 FineTranslations

See the tables above for the subset of the language you want to download.

We currently do not provide smaller sample versions, but by setting limit or using streaming=True you can easily fetch a sample of the data.

Using 🏭 datatrove

from datatrove.pipeline.readers import ParquetReader

# limit determines how many documents will be streamed (remove for all)
# this will fetch the Portuguese filtered data
data_reader = ParquetReader("hf://datasets/HuggingFaceFW/finetranslations/data/por_Latn", limit=1000)
for document in data_reader():
    # do something with document
    print(document)

###############################    
# OR for a processing pipeline:
###############################

from datatrove.executor import LocalPipelineExecutor
from datatrove.pipeline.readers import ParquetReader
from datatrove.pipeline.filters import LambdaFilter
from datatrove.pipeline.writers import JsonlWriter

pipeline_exec = LocalPipelineExecutor(
    pipeline=[
        ParquetReader("hf://datasets/HuggingFaceFW/finetranslations/data/por_Latn", limit=1000),
        LambdaFilter(lambda doc: "hugging" in doc.text),
        JsonlWriter("some-output-path")
    ],
    tasks=10
)
pipeline_exec.run()

Using huggingface_hub

from huggingface_hub import snapshot_download
folder = snapshot_download(
                "HuggingFaceFW/finetranslations", 
                repo_type="dataset",
                local_dir="./finetranslations/",
                # download the Czech filtered data
                allow_patterns=["data/ces_Latn/*"])

Using datasets

from datasets import load_dataset
# get data from all languages
fw = load_dataset("HuggingFaceFW/finetranslations", name="all", split="train", streaming=True)

Dataset processing steps

We used the 🏭 datatrove library to process the data. You can find the entire working code that created the dataset here.

1. Sourcing the data

The starting point for this dataset was our previously released 🥂 FineWeb2 dataset, a large scale pre-training dataset covering over a thousand languages.

As many of these language subsets consisted in large part of religious content (mostly bibles) or Wikipedia pages, we only included the languages whose subset had a bible_wiki_ratio (ratio of documents with this type of content) under 0.5 (around 500 languages).

We processed up to 50B tokens per language. For languages that originally had more than 50B tokens, we employed quality classifiers from FineWeb2-HQ and kept the top 50B tokens. When a classifier wasn't available, we randomly sampled 50B tokens worth of documents.

2. Running translation at scale

We compared a variety of models from the Qwen, Llama, Gemma, Mistral and Aya models on translation benchmarks on a large number of languages. Qwen and Gemma models showed the strongest performance across the board, but Qwen models would sometimes output Chinese even when translating from European languages into English. As such, and for simplicity, we employed Gemma3 27B to translate all languages.

The main issues we observed from our early experiments were:

• a large amount of toxic/adult/gambling related content, specially originating from our lower resource languages;
• lack of adherence to the original formatting. In particular, new lines would often be removed or added arbitrarily;
• repetition loops (that would run until the model context was full), particularly in very large documents

We relied on the following measures to address them:

• have the model initially classify the type of content before translating, flagging adult/spam like content early (faster processing too)
• strict formatting rules in the prompt
• chunk documents into at most 512 token chunks. We then rely on a sliding window approach to translate the next chunk while keeping the previous one (already translated) in the prompt for context

The prompt used is as follows:

Click to view the full translation SYSTEM prompt

**EARLY EXIT (runs BEFORE anything else)**
1) First, classify <ORIGINAL>.
2) If it contains ANY of:
   - Pornographic/explicit sexual content (incl. escorting)
   - Online gambling/casino/betting
   - Trading/crypto/forex promotional content
   - Lists of unrelated keywords or phrases lacking complete sentences and grammatical connectors (SEO spam)
THEN immediately output exactly:
<TRANSLATION>CONTENT FLAG</TRANSLATION>
and STOP — ignore all other instructions.
3) If not, proceed with translation rules.

You are a professional translator. Follow **all** instructions exactly.

**Crucial Formatting Rules (READ CAREFULLY — HARD REQUIREMENTS):**
1. **Preserve formatting EXACTLY.**
* Do **not** add, remove, or modify any line breaks.
* You must output the **exact same number of lines** as the input.
* Each line in your output must correspond exactly to one line in the input.
* You must **never** insert additional blank lines that do not exist in the original text.
* Do **not** insert blank lines for readability.
2. **Translate every token.**
Do not skip, summarize, or ignore any word, punctuation mark, or spacing.
3. **No literal translation.**
Make the English natural and fluent, but **do not** change formatting.
4. **No hallucinations.**
Do not add explanations, commentary, or any content that isn't in the original.
5. **Output format**
Enclose the translated text **only** in:
<TRANSLATION>
</TRANSLATION>
Nothing before or after.
6. **If you cannot follow these formatting rules exactly, output:**
`ERROR: formatting rule violated`

**Additional Strict Requirements:**
- Do **NOT** insert extra whitespace.
- Do **NOT** auto-format paragraphs.
- Do **NOT** add blank lines.
- Do NOT reinterpret or restructure the text. Do NOT treat long lines as paragraphs. You must preserve every line exactly as written, even if the line is extremely long, contains many sentences, or appears to represent multiple paragraphs.
- Do NOT split any lines into multiple lines. Even if a line contains many sentences, you must keep it as a single line exactly as in the original.
- Treat every visible line break as unchangeable.
- When in doubt, copy the structure line by line.

Remember the early exit rule before you consider translating.

Click to view the full translation USER prompt

**{display_language} ({subset_language}) Text to Translate (preserve all line breaks EXACTLY):**
<ORIGINAL>{combined_chunk}</ORIGINAL>

Now translate to English (eng_Latn).

The pipeline ran on the Hugging Face cluster over a period of 3 months, making use of spare compute cycles.

3. Post-processing

We removed content that the model had flagged, removed the <TRANSLATION></TRANSLATION> markdown tags and ensured line breaks on the chunk boundaries were consistent.

4. Edu-filtering

We attempted to use the quality classifier from FineWeb-Edu to boost English performance. However, this filtering did not lead to a performance improvement, potentially due to distribution differences between Gemma generated text and natural English from web pages. Therefore, we trained a new classifier that provided a modest performance boost when filtering for the top 10% of content. This content is available in the finetranslations-edu dataset.

Dataset card for 💬 FineTranslations

Dataset Summary

This dataset contains over 1 trillion tokens of parallel text in English and 500+ languages. It was obtained by translating data from 🥂 FineWeb2 into English using Gemma3 27B.

Dataset Structure

Data Instances

The following is an example sample from the dataset. It is part of the French (fra_Latn) data, originally from the CC-MAIN-2014-52 CommonCrawl snapshot.

{
   "translated_text": "A and I completed a small DIY project that is easy but yields surprising and charming results.\nWe used wooden letters that A patiently painted with wood paint.\nSubsequently, we added buttons with hot glue to embellish everything.\nEach person now has their own colorful and fun letter!",
   "translated_chunks": [
      "A and I completed a small DIY project that is easy but yields surprising and charming results.\nWe used wooden letters that A patiently painted with wood paint.\nSubsequently, we added buttons with hot glue to embellish everything.\nEach person now has their own colorful and fun letter!"
   ],
   "og_chunks": [
      "A et moi avons réalisé un petit projet brico facile mais qui donne des résultats surprenants et charmants.\nNous avons utilisé des lettres de bois que A a patiemment peint avec de la peinture à bois.\nPar la suite, nous y avons ajouté, à la colle chaude, des boutons pour garnir le tout.\nChacun a maintenant sa propre lettre colorée et amusante!"
   ],
   "og_full_text": "A et moi avons réalisé un petit projet brico facile mais qui donne des résultats surprenants et charmants.\nNous avons utilisé des lettres de bois que A a patiemment peint avec de la peinture à bois.\nPar la suite, nous y avons ajouté, à la colle chaude, des boutons pour garnir le tout.\nChacun a maintenant sa propre lettre colorée et amusante!",
   "og_language": "fra_Latn",
   "og_language_score": 0.9992175698280334,
   "og_token_count": 83,
   "og_quality_score": 0.03385915607213974,
   "early_stop": false,
   "id": "<urn:uuid:d7835f7d-d5e5-451e-97fb-6d51bf8addcf>",
   "url": "http://mcommemaman.blogspot.com/2008/12/bricolage-personnalis.html",
   "warc_path": "s3://commoncrawl/crawl-data/CC-MAIN-2014-52/segments/1418802778085.5/warc/CC-MAIN-20141217075258-00136-ip-10-231-17-201.ec2.internal.warc.gz",
   "minhash_cluster_size": 42,
   "translated_token_count": 58,
   "edu_score_raw": 0.640625,
   "edu_score": 1
}

Data Fields

• translated_text (string): the English translated text content (all chunks concatenated)
• translated_chunks (list of strings): the English translation split into chunks
• og_chunks (list of strings): the original text in the source language split into chunks (there is a 1-1 match between translated_chunks and og_chunks)
• og_full_text (string): the original full text in the source language
• og_language (string): language-script code for the original text (e.g., fra_Latn)
• og_language_score (float): language prediction score for the original text as reported by the GlotLID classifier
• og_token_count (int): token count of the original text
• og_quality_score (float): quality score of the original text from the EPFL classifiers (if available, -1 otherwise)
• early_stop (bool): whether translation stopped early due to formatting issues. In this case, the last chunks from the original text were dropped.
• id (string): original unique identifier for this sample from CommonCrawl
• url (string): url to the original page where the text was present
• warc_path (string): s3 path for the individual CommonCrawl warc file containing this sample
• minhash_cluster_size (int): number of samples in the FineWeb2 minhash cluster of this sample. See the deduplication section of FineWeb2 for more info.
• translated_token_count (int): token count of the English (translated) text
• edu_score_raw (float): raw educational score from the educational classifier
• edu_score (int): binned educational score (0-4 scale)

Notes:

• og_quality_score is from a classifier applied on the original language, while edu_score was computed on the translated English text
• in case of early_stop, some chunks might have been dropped. In this case, og_full_text might not match translated_text. Rely on the chunk variables if needed.

Data Splits

See "Languages and available subsets" above.

Dataset Creation

Curation Rationale

The main motivation behind the creation of this dataset was improving translation capabilities. While models are generally strong at translating from other languages into English (X->English), the opposite is often not true, particularly for lower resource languages. Our approach was to take data that was originally in non-English languages (from 🥂 FineWeb2, our large multilingual pre-training dataset), chunk it, and translate it. This dataset can then be used to improve English->X translations by finetuning existing models.

Additionally, the resulting English data contains relevant cultural information for different countries and languages, and our experiments show that the 1T tokens we obtained perform on a similar level as our 🍷 FineWeb dataset. This data can therefore also be used for English only model training (potentially as an extension of FineWeb).

Source Data

The source data for 💬 FineTranslations is 🥂 FineWeb2, a large scale pre-training dataset covering over a thousand languages sourced from CommonCrawl webpages crawled over the 2013-2024 time period.

As many of these language subsets consisted in large part of religious content (mostly bibles) or Wikipedia pages, we only included the languages whose subset had a bible_wiki_ratio (ratio of documents with this type of content) under 0.5 (around 500 languages).

We processed up to 50B tokens per language. For languages that originally had more than 50B tokens, we employed quality classifiers from FineWeb2-HQ and kept the top 50B tokens. When a classifier wasn't available, we randomly sampled 50B tokens worth of documents.

Data processing steps

See "Dataset processing steps" above.

Annotations

We augment the original samples with translation-related annotations including translated_text, translated_chunks, og_chunks, og_full_text, og_language, og_language_score, og_token_count, og_quality_score, translated_token_count, early_stop, edu_score_raw, and edu_score. The original language annotations (og_language, og_language_score) are inherited from FineWeb2 and were automatically generated by the language filter. The minhash_cluster_size is also inherited from FineWeb2 and was computed during the deduplication process. Translation-specific annotations track the translation process, quality, and educational scores.

Personal and Sensitive Information and opt-out

The source data (FineWeb2) anonymizes email addresses and public IP addresses.

For emails, a regex pattern is applied and any occurrence of an email address is replaced with either [email protected] or [email protected]. For IP addresses, a regex pattern is employed and then further filtered to only anonymize IP addresses allocated for public networks. Matched IP addresses are then replaced with one of the following randomly generated IP addresses, which at the time of dataset creation were not responding to ping requests: 22.214.171.124, 126.96.36.199, 188.8.131.52, 184.108.40.206, 220.127.116.11, and 18.104.22.168. The source dataset decided against applying regex patterns for phone numbers due to the high false positive rate.

Despite these efforts, given that 💬 FineTranslations is sourced from web content at large, it is very likely that some personally identifiable information (PII) will be present. If you find your own PII in 💬 FineTranslations and would like it removed, please fill out our PII removal/opt out form.

CommonCrawl respects robots.txt at crawl time, but if you are a webmaster and find your website in 💬 FineTranslations and would like to have it removed, you may also use the PII removal/opt out form.

Considerations for Using the Data

Social Impact of Dataset

With the release of this dataset we aim to improve translation capabilities, particularly for lower resource languages where English->X translation is often weak. By providing over 1 trillion tokens of parallel text data across 500+ languages, we enable researchers and practitioners to:

• Improve translation models: Finetune existing models on this parallel data to improve English->X translation capabilities
• Train multilingual models: Use the parallel data for training or improving multilingual models
• Enhance English models: Leverage the translated English content, which contains cultural information from diverse languages and performs similarly to FineWeb for English-only training

The dataset is fully reproducible with all code available, making the translation pipeline transparent and allowing the community to build upon our work.

Discussion of Biases

The dataset inherits biases from both the source data (FineWeb2) and the translation process:

Source data biases: As FineWeb2 was sourced from the web, any harmful biases typically present in web content may be reproduced in this dataset. Efforts were made in the source dataset to minimize NSFW and toxic content through URL-level filtering, but some toxic or harmful content may still be present.

Translation model biases: The translations were generated using Gemma3 27B, which may introduce its own biases:

• The model may translate certain concepts or cultural references in ways that don't fully capture the original meaning
• Translation quality may vary across languages, with lower resource languages potentially receiving lower quality translations
• The model's training data biases may be reflected in the translations

Content filtering: We employed early exit mechanisms to flag adult/spam content before translation, but some content that passed these filters may still be considered inappropriate. The formatting preservation requirements may also have led to some translations that don't read as naturally as human translations.

Other Known Limitations

Translation quality: While we compared multiple models and selected Gemma3 27B for its strong performance, translation quality is not uniform across all languages. Lower resource languages may have lower translation quality, and some translations may contain errors or awkward phrasing.

Formatting preservation: We prompt the translation model to strictly maintain the original formatting, such as line breaks and document structure. However, in practice, the model does not always fully respect these instructions—so while our approach aims for high formatting fidelity, there may be cases where formatting inconsistencies remain or the structure is not perfectly preserved.

Model limitations: The translation model has context limitations (we chunked documents into 512 token chunks), which means very long documents are translated in pieces. While we use a sliding window approach to maintain context, some coherence may be lost across chunk boundaries.

Language coverage: We only included languages from FineWeb2 with a bible_wiki_ratio under 0.5, which excluded some languages that were predominantly religious or Wikipedia content. This means the dataset may not be representative of all possible language content.

Educational filtering: The educational classifier was trained specifically for this dataset, but its performance may vary across different types of content. The top 10% educational content is available in a separate dataset (finetranslations-edu).

We encourage users to review the translation quality for their languages of interest and consider additional filtering or post-processing if needed.

Additional Information

Licensing Information

The dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0 license. The use of this dataset is also subject to CommonCrawl's Terms of Use.

Citation Information

@misc{penedo2026finetranslations,
      title={FineTranslations}, 
      author={Guilherme Penedo and Hynek Kydl{\'\i}{\v{c}}ek and Amir Hossein Kargaran and Leandro von Werra},
      year={2026},
      publisher = {Hugging Face},
      journal = {Hugging Face repository},
      howpublished = {\url{https://huggingface.co/datasets/HuggingFaceFW/finetranslations}}
}