Create README.md

README.md

@@ -0,0 +1,77 @@
+---
+task_categories:
+- translation
+- text-classification
+language:
+- en
+- fr
+tags:
+- English
+- French
+- smt
+size_categories:
+- 1M<n<10M
+---
+# English-French Translation Dataset
+
+This dataset provides a large collection of English-French translation pairs. It was downloaded from the [OPUS collection](https://opus.nlpl.eu/) on 2025-02-26.
+
+These datasets cover basic French, news, presentations (like TED talks and academic conferences), and educational videos. This helps avoid the bias towards legal, administrative, and bureaucratic content found in datasets like Europarl and ECB.
+
+## Data Processing
+
+*   **Normalization:** Bullet points, hyphenated words, quotation marks, Unicode characters, and whitespace have been normalized.
+*   **Placeholder Replacement:**  E-mails, emojis, hashtags, phone numbers, URLs, and user handles have been replaced with placeholders to protect privacy and maintain data integrity.
+*   **Deduplication:** Duplicate (English, French) pairs have been removed.
+
+## Dataset Structure
+
+*   `raw.jsonl`:  Contains all deduplicated translation pairs in JSON Lines format.
+*   `train.jsonl`: File containing the training split (80% of the filtered data).
+*   `test.jsonl`: File containing the test split (20% of the filtered data).
+
+The training and test splits were created after filtering for translation quality scores > 0.8 and stratified by the source dataset to ensure representation across different sources.
+
+## Data Fields
+
+Each row in the JSONL files is a JSON object with the following fields:
+
+*   `id`:  A unique identifier for each translation pair.
+*   `english`:  The English text.
+*   `french`: The French translation.
+*   `source`: The original dataset from which the pair was extracted.
+*   `translation_quality`: A score representing the quality of the translation, calculated using LaBSE cosine similarity. Higher values indicate better translation quality.
+*   `readability_grade`: A readability score assessed using [agentlans/deberta-v3-base-zyda-2-readability](agentlans/deberta-v3-base-zyda-2-readability) on the English text.  This indicates the grade level required to understand the text.
+
+### Example Row
+
+Note that the French non-ASCII characters have been escaped:
+```
+{
+  "id": 625153,
+  "english": "And we also sampled the outdoor air.",
+  "french": "Enfin, nous avons aussi pr\\u00e9lev\\u00e9 l'air ext\\u00e9rieur.",
+  "source": "TED2020",
+  "translation_quality": 0.8045,
+  "readability_grade": 3.12
+}
+```
+
+## Dataset Statistics
+
+The following table shows the class distribution across the training and test sets:
+
+| Source                     | Train    | Test   | Total    |
+| -------------------------- | --------:| ------:| --------:|
+| ELRC-1118-CORDIS\_News     | 97866    | 24467  | 122333   |
+| ELRC-5067-SciPar           | 626727   | 156682 | 783409   |
+| GlobalVoices               | 119330   | 29832  | 149162   |
+| News-Commentary            | 90026    | 22507  | 112533   |
+| QED                        | 301807   | 75452  | 377259   |
+| TED2020                    | 249578   | 62395  | 311973   |
+| Tatoeba                    | 201522   | 50380  | 251902   |
+| **Total**                  | **1686856** | **421715** | **2108571** |
+
+## Intended Uses
+
+This dataset is suitable for training and evaluating machine translation models, as well as for research in natural language processing, cross-lingual understanding, and related fields.  The `translation_quality` and `readability_grade` fields can be used for filtering and analysis.