Update README.md

README.md

@@ -1,99 +1,119 @@
 ---
-configs:
-- config_name: ara
-  data_files:
-  - split: train
-    path: data/ara/train_marginmse.jsonl
-- config_name: dan
-  data_files:
-  - split: train
-    path: data/dan/train_marginmse.jsonl
-- config_name: deu
-  data_files:
-  - split: train
-    path: data/deu/train_marginmse.jsonl
-- config_name: eng
-  data_files:
-  - split: train
-    path: data/eng/train_marginmse.jsonl
-- config_name: fas
-  data_files:
-  - split: train
-    path: data/fas/train_marginmse.jsonl
-- config_name: fra
-  data_files:
-  - split: train
-    path: data/fra/train_marginmse.jsonl
-- config_name: hin
-  data_files:
-  - split: train
-    path: data/hin/train_marginmse.jsonl
-- config_name: ind
-  data_files:
-  - split: train
-    path: data/ind/train_marginmse.jsonl
-- config_name: ita
-  data_files:
-  - split: train
-    path: data/ita/train_marginmse.jsonl
-- config_name: jpn
-  data_files:
-  - split: train
-    path: data/jpn/train_marginmse.jsonl
-- config_name: kor
-  data_files:
-  - split: train
-    path: data/kor/train_marginmse.jsonl
-- config_name: nld
-  data_files:
-  - split: train
-    path: data/nld/train_marginmse.jsonl
-- config_name: pol
-  data_files:
-  - split: train
-    path: data/pol/train_marginmse.jsonl
-- config_name: por
-  data_files:
-  - split: train
-    path: data/por/train_marginmse.jsonl
-- config_name: rus
-  data_files:
-  - split: train
-    path: data/rus/train_marginmse.jsonl
-- config_name: spa
-  data_files:
-  - split: train
-    path: data/spa/train_marginmse.jsonl
-- config_name: swe
-  data_files:
-  - split: train
-    path: data/swe/train_marginmse.jsonl
-- config_name: tur
-  data_files:
-  - split: train
-    path: data/tur/train_marginmse.jsonl
-- config_name: vie
-  data_files:
-  - split: train
-    path: data/vie/train_marginmse.jsonl
-- config_name: zho
-  data_files:
-  - split: train
-    path: data/zho/train_marginmse.jsonl
-
+language:
+- ara
+- dan
+- deu
+- eng
+- fas
+- fra
+- hin
+- ind
+- ita
+- jpn
+- kor
+- nld
+- pol
+- por
+- rus
+- spa
+- swe
+- tur
+- vie
+- zho
+multilingual: true
+tags:
+- dense-retrieval
+- hard-negatives
+- knowledge-distillation
+- webfaq
+license: cc-by-4.0
+task_categories:
+- sentence-similarity
+- text-retrieval
+size_categories:
+- 1M<n<10M
 ---
 
-# WebFAQ Distilled (MarginMSE)
+# WebFAQ 2.0: Multilingual Hard Negatives
+
+This dataset contains **mined hard negatives** derived from the **WebFAQ 2.0** corpus. It includes approximately **1.3 million** samples across **20 languages**.
+
+The dataset is designed to support robust training of dense retrieval models, specifically enabling:
+1.  **Contrastive Learning:** Using strict hard negatives to improve discrimination.
+2.  **Knowledge Distillation:** Using the provided cross-encoder scores to train with soft labels (e.g., MarginMSE).
+
+## Dataset Creation & Mining Process
+
+To ensure high-quality training signals, we employed a **two-stage mining pipeline** that balances difficulty with correctness.
+
+### 1. Lexical Retrieval (Recall)
+For every query in WebFAQ, we first retrieved the **top-200 candidate answers** from the monolingual corpus using **BM25**.
+* **Goal:** Identify candidates with high lexical overlap (shared keywords) that are likely to be "hard" for a dense retriever to distinguish.
+
+### 2. Semantic Reranking (Precision)
+We reranked the top-200 candidates using the state-of-the-art cross-encoder model: **[BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3)**.
+* **Goal:** Assess the true semantic relevance of each candidate.
+
+### 3. Filtering & Scoring
+We applied a rigorous filtering strategy to curate the final dataset:
+* **False Negative Removal:** Candidates with extremely high cross-encoder scores (semantic matches) were discarded to prevent "poisoning" the training data with valid answers labeled as negatives.
+* **Easy Negative Removal:** Candidates with very low scores were discarded to ensure training efficiency.
+* **Score Retention:** We retained the BGE-M3 relevance scores for every negative, enabling knowledge distillation workflows.
+
+## Dataset Structure
+
+The data is stored in a **grouped format** (JSONL), where each line represents a single query paired with its positive answer and a list of mined hard negatives.
+
+Each sample contains:
+
+| Field | Type | Description |
+| :--- | :--- | :--- |
+| `query` | String | The user question. |
+| `positive` | String | The ground-truth correct answer. |
+| `positive_score` | Float | The **BGE-M3** relevance score for the positive answer. |
+| `negatives` | List[String] | A list of mined hard negatives (non-relevant but similar). |
+| `negative_scores`| List[Float] | The **BGE-M3** relevance scores corresponding to each negative. |
+
+**Note:** This format is optimized for:
+* **Contrastive Learning:** You can instantly sample 1 positive and $N$ negatives.
+* **MarginMSE:** You have all the teacher scores (positive and negative) required to compute the margin loss.
+
+## Languages & Distribution
+
+The dataset covers **20 languages** with the following sample counts:
+
+| ISO Code | Language | Samples |
+| :--- | :--- | :--- |
+| `ara` | Arabic | 32,000 |
+| `dan` | Danish | 32,000 |
+| `deu` | German | 96,000 |
+| `eng` | English | 128,000 |
+| `fas` | Persian | 64,000 |
+| `fra` | French | 96,000 |
+| `hin` | Hindi | 32,000 |
+| `ind` | Indonesian | 32,000 |
+| `ita` | Italian | 96,000 |
+| `jpn` | Japanese | 96,000 |
+| `kor` | Korean | 32,000 |
+| `nld` | Dutch | 96,000 |
+| `pol` | Polish | 64,000 |
+| `por` | Portuguese | 64,000 |
+| `rus` | Russian | 96,000 |
+| `spa` | Spanish | 96,000 |
+| `swe` | Swedish | 32,000 |
+| `tur` | Turkish | 32,000 |
+| `vie` | Vietnamese | 32,000 |
+| `zho` | Chinese | 32,000 |
+| **Total** | **All** | **~1,280,000** |
 
-This dataset contains **Hard Negatives** and **Teacher Scores** (from BGE-M3) for the WebFAQ dataset.
-It is designed for training dense retrievers using **MarginMSE Loss**.
+## Citation
 
-## Structure
-- **query**: The user question.
-- **positive**: The correct answer.
-- **positive_score**: The teacher's score for the positive pair.
-- **negatives**: A list of hard negatives mined via BM25.
-- **negative_scores**: The teacher's scores for each negative.
+If you use this dataset, please cite the WebFAQ 2.0 paper:
 
-## Languages
-20 languages included.
+```bibtex
+@inproceedings{dinzinger2025webfaq,
+  title={WebFAQ: A Multilingual Collection of Natural QA Datasets for Dense Retrieval},
+  author={Dinzinger, Michael and Caspari, Laura and Dastidar, Kanishka Ghosh and Mitrović, Jelena and Granitzer, Michael},
+  booktitle={Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
+  year={2025}
+}