Update README.md

README.md

@@ -1,86 +1,80 @@
 ---
 language:
-  - th
-license: apache-2.0
+- th
+license: other
 task_categories:
-  - text-generation
-  - fill-mask
-tags:
-  - thai
-  - pretraining
-  - deduplicated
-  - minhash
-  - consensus
-pretty_name: ThaiMix
-size_categories:
-  - 10M<n<100M
+- text-generation
+arxiv: 2512.18834
 configs:
-  - config_name: consensus
-    data_files:
-      - split: train
-        path: consensus/*.parquet
-  - config_name: minhash_deduped
-    data_files:
-      - split: train
-        path: minhash_deduped/*.parquet
-default: true
+- config_name: minhash_deduped
+  data_files:
+  - split: train
+    path: minhash_deduped/*.parquet
+- config_name: matched
+  data_files:
+  - split: train
+    path: consensus/*.parquet
+default: minhash_deduped
 ---
 
-# ThaiMix
+<p align="center">
+  <a href="https://huggingface.co/collections/AdaMLLab/mixminmatch">
+    <img src="https://img.shields.io/badge/🤗_Collection-MixMinMatch-blue" alt="MixMinMatch Collection">
+  </a>
+</p>
 
-A high-quality Thai language pretraining dataset created by combining and deduplicating multiple sources.
+ThaiMix ([https://arxiv.org/abs/2512.18834](https://arxiv.org/abs/2512.18834)) is a Thai pretraining corpus containing 70 billion tokens across 81 million documents (in the minhash subset). Rather than scraping the web again, ThaiMix combines five publicly available Thai datasets, applies Thai-specific quality filtering, and performs cross-dataset deduplication.
 
-## Dataset Description
+## Subsets
 
-This dataset contains Thai text collected from multiple web sources, deduplicated using MinHash (80% Jaccard similarity threshold), and filtered for quality.
+| Subset | Documents | Tokens | Description |
+|--------|-----------|--------|-------------|
+| `minhash_deduped` | 81.3M | 70.5B | Document-level MinHash deduplication |
+| `matched` | 10.9M | 12.3B | Documents appearing in 2+ source datasets |
 
-### Subsets
-
-| Subset | Description | Documents | Schema |
-|--------|-------------|-----------|--------|
-| `consensus` | Documents appearing in 2+ sources (high quality) | ~10.9M | text, id, source (list) |
-| `minhash_deduped` | Full deduplicated dataset | ~68M | text, id, source |
-
-### Sources
-
-The dataset combines text from:
-- **C4** - Colossal Clean Crawled Corpus
-- **CulturaX** - Multilingual web corpus
-- **FineWeb2** - High-quality web text
-- **HPLT2** - High Performance Language Technologies
-- **SEA-CC** - Southeast Asian Common Crawl
+The matched subset uses cross-dataset agreement as a signal for quality.
 
 ## Usage
 
 ```python
 from datasets import load_dataset
 
-# Load consensus subset (recommended for high-quality pretraining)
-ds = load_dataset("AdaMLLab/ThaiMix", "consensus")
-
-# Load full deduplicated dataset
 ds = load_dataset("AdaMLLab/ThaiMix", "minhash_deduped")
+ds = load_dataset("AdaMLLab/ThaiMix", "matched")
 ```
 
-## Schema
+## Sources
 
-### Consensus subset
-- `text` (string): The document text
-- `id` (string): Unique document identifier (MD5 hash)
-- `source` (list[string]): List of sources where this document appears
+Tokens were counted using `meta-llama/Llama-3.2-3B`'s tokenizer.
 
-### Minhash_deduped subset
-- `text` (string): The document text
-- `id` (string): Unique document identifier
-- `source` (string): Original source of the document
+| Source | Survival Rate |
+|--------|---------------|
+| HPLT 2.0 | 75.1% |
+| C4 | 72.6% |
+| SEA-CC | 70.7% |
+| CulturaX | 69.0% |
+| FineWeb-2 | 61.7% |
 
-## Processing Pipeline
+## Pipeline
 
-1. **Download**: Raw data from each source
-2. **Filtering**: Language detection, quality filters (character ratios, stop words, etc.)
-3. **MinHash Deduplication**: 80% Jaccard similarity threshold with 128 hash functions
-4. **Consensus Building**: Identify documents appearing in 2+ sources using MinHash clusters
+1. Quality filtering with Thai-specific thresholds (Thai script ratio, repetition patterns, language identification)
+2. Document-level MinHash deduplication (5-gram shingles, 14 bands, 8 hashes per band, similarity threshold 0.8)
+3. Cross-source matching to identify documents appearing in 2+ independent sources
+
+## Citation
+
+```bib
+@misc{alrashed2025mixminmatch,
+      title={Mix, MinHash, and Match: Cross-Source Agreement for Multilingual Pretraining Datasets}, 
+      author={Sultan Alrashed and Francesco Orabona},
+      year={2025},
+      eprint={2512.18834v2},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2512.18834v2}, 
+}
+```
 
 ## License
 
-Apache 2.0
+See individual source dataset licenses.