Datasets:
|
Download README.md from samdei/samdei-type-pretraining-data: direct link, hf CLI and curl.
- Browser
- Download file 10.4 kB
-
https://huggingface.co/datasets/samdei/samdei-type-pretraining-data/resolve/main/README.md
- Command line
-
hf download hf://datasets/samdei/samdei-type-pretraining-data/README.md
-
curl -L -o README.md https://huggingface.co/datasets/samdei/samdei-type-pretraining-data/resolve/main/README.md
10.4 kB
| pretty_name: Samdei Type Pretraining Data | |
| license: cc-by-sa-4.0 | |
| language: | |
| - km | |
| task_categories: | |
| - translation | |
| tags: | |
| - transliteration | |
| - romanization | |
| - romanized-khmer | |
| - khmer | |
| - synthetic | |
| size_categories: | |
| - 10M<n<100M | |
| multilinguality: monolingual | |
| source_datasets: | |
| - extended|HPLT/HPLT2.0_cleaned | |
| - extended|HuggingFaceFW/fineweb-2 | |
| - extended|wikimedia/wikipedia | |
| - extended # + ALT (Asian Language Treebank, NICT), downloaded from NICT, not from the Hub | |
| annotations_creators: | |
| - machine-generated | |
| language_creators: | |
| - found | |
| - machine-generated | |
| configs: | |
| - config_name: generator_v2 | |
| default: true | |
| data_files: | |
| - split: train | |
| path: generator_v2/train_v2_s3.parquet | |
| - split: validation | |
| path: generator_v2/train_v2_s3.dev.parquet | |
| - config_name: generator_v1 | |
| data_files: | |
| - split: train | |
| path: | |
| - generator_v1/train_web_s1.parquet | |
| - generator_v1/train_web_s2.parquet | |
| - split: validation | |
| path: generator_v1/train_web_s1.dev.parquet | |
| - config_name: llm_chat_sentences | |
| data_files: | |
| - split: train | |
| path: llm/train_gemma_s1.parquet | |
| - config_name: llm_romanized | |
| data_files: | |
| - split: train | |
| path: llm/train_gemrom.parquet | |
| # Samdei Type Pretraining Data | |
| These are the exact **pretraining (stage-1)** pairs of the Samdei Type romanized-Khmer → Khmer-script converters: 19.3M (romanized input, Khmer target) pairs. Every **Khmer target is real text**: 1.41M sentences from HPLT v2, FineWeb-2, Khmer Wikipedia and ALT, plus 2,483 chat sentences written by Gemma 4. The **romanized input** is made by a rule-based generator that spells each piece of a sentence as one of four typing personas would type it, and for a small part by Gemma 4. | |
| They trained the pretrained checkpoints [`samdei/samdei-type-base-pretrained`](https://huggingface.co/samdei/samdei-type-base-pretrained) and [`samdei/samdei-type-mini-pretrained`](https://huggingface.co/samdei/samdei-type-mini-pretrained), from which Samdei Type Base and Mini were fine-tuned. The fine-tuning and adaptation data (stages 2–3) is in [`samdei/samdei-type-finetuning-data`](https://huggingface.co/datasets/samdei/samdei-type-finetuning-data). | |
| Paper: [technical report](https://github.com/Samdei-Labs/samdei-type/blob/main/paper/samdei-type.pdf). Inference code: [github.com/Samdei-Labs/samdei-type](https://github.com/Samdei-Labs/samdei-type). Test set: [`samdei/samdei-type-bench`](https://huggingface.co/datasets/samdei/samdei-type-bench). | |
| ## Configs and files | |
| The data is organized by what it is, not by model. `recipes.json` gives the exact mix of each model (files, repeats, steps, learning rate), read from the training arguments stored in the checkpoints. | |
| | Config | Files (rows) | What it is | Used by | | |
| |---|---|---|---| | |
| | `generator_v2` (default) | `train_v2_s3` (6,501,912) | web sentences and the chat sentences, spelled by the second generator version | Mini | | |
| | `generator_v1` | `train_web_s1` (6,289,947), `train_web_s2` (6,288,484) | web sentences, spelled twice by the first generator version (two independent samples) | Base; the channel model (`train_web_s1`) | | |
| | `llm_chat_sentences` | `train_gemma_s1` (189,431) | the 2.4k LLM-written chat sentences, spelled by the first generator version, ×40 | Base; the channel model | | |
| | `llm_romanized` | `train_gemrom` (5,715) | Gemma 4 romanizations of 2,000 real sentences (×3 copies with random casing) | Mini (listed ×3 in its mix); the channel model | | |
| | Model | Stage-1 mix | Pairs | | |
| |---|---|---| | |
| | Base (18.2M) | `train_web_s1` + `train_web_s2` + `train_gemma_s1` | 12,767,862 | | |
| | Mini (7.1M) | `train_v2_s3` + `train_gemrom` ×3 | 6,519,057 | | |
| | channel model (Khmer → Latin, used in Base's reranker) | `train_web_s1` + `train_gemma_s1` + `train_gemrom` | 6,485,093 | | |
| **Why two generator versions.** Base was pretrained first, with the first generator version. The second version adds: | |
| - English chat words and place names inserted on both sides; | |
| - standalone place-name pairs (`persona` = `place`); | |
| - sentence-final punctuation dropped from both sides together. | |
| It came out of the errors of the first models. Mini was pretrained with it, and both models' later stages use it. Both versions are here, so both models' pretraining can be reproduced exactly. | |
| **Validation splits** are synthetic: held-out sentences spelled by the same generator, monitored during training. Checkpoints were selected on separate evaluation sets, which are not part of this dataset. | |
| ``` | |
| generator_v1/ train_web_s1.parquet train_web_s2.parquet train_web_s1.dev.parquet | |
| generator_v2/ train_v2_s3.parquet train_v2_s3.dev.parquet | |
| llm/ train_gemma_s1.parquet train_gemrom.parquet | |
| recipes.json | |
| ``` | |
| ## Columns | |
| | Column | Content | | |
| |---|---| | |
| | `latin` | romanized input, as the model saw it | | |
| | `khmer` | Khmer-script target, normalized with `khnormal`; English words, digits and brand names are copied from the input | | |
| | `persona` | `formal`, `careful`, `mixed` or `lazy` (the generator's typing personas, from standard spelling to heavy chat shortcuts and typos); `place` (gazetteer pair); `gemrom0`–`gemrom3` (Gemma 4 with one of four style prompts) | | |
| | `source` | where the Khmer target comes from: `hplt`, `fineweb2`, `wikipedia`, `alt` (a piece of a web sentence); `gemma` (a piece of an LLM-written chat sentence); `gazetteer`; `gemrom_<x>` (a whole sentence from `<x>`, romanized by Gemma 4) | | |
| ## How to load | |
| ```python | |
| from datasets import load_dataset | |
| ds = load_dataset("samdei/samdei-type-pretraining-data", "generator_v2", split="train") | |
| print(ds[0]) # {'latin': ..., 'khmer': ..., 'persona': ..., 'source': ...} | |
| ``` | |
| To rebuild a mix exactly, read `recipes.json` and concatenate its files in order, with repeats. | |
| ## How the pairs were made | |
| 1. **Khmer targets.** | |
| - 1.41M web sentences: HPLT v2 802k (3 of 13 shards of khm_Khmr), FineWeb-2 280k, Khmer Wikipedia 306k (20231101 dump) and the Khmer side of ALT 22k. | |
| - Each sentence is normalized with `khnormal`, split and deduplicated. | |
| - Lines with URLs or e-mail-like strings, and lines that are mostly Latin, are dropped. | |
| - Sentences that overlap an evaluation set are removed before any pair is made. | |
| - 2,483 chat sentences were written by Gemma 4 12B, run locally, across 50 chat domains and 4 registers. | |
| 2. **The persona generator:** | |
| 1. It cuts each sentence into pieces of 1–12 words (mostly 5–6) with `khmercut`. | |
| 2. It pronounces each word with Google's Khmer pronunciation lexicon (62k words, CC BY 4.0); rule-based G2P covers the rest. | |
| 3. It spells the piece in a sampled persona: formal 8%, careful 32%, mixed 40%, lazy 20%. Personas differ in spelling choices (`nh / ny / gn`, final consonants, vowels) and in how often they use a hand-written chat lexicon (`knhom`, `kyom` or `nh` for ខ្ញុំ). | |
| 4. It adds noise: glued or split words, dropped vowels, k/g and t/d swaps, mistyped keys. | |
| 3. **LLM romanizations.** Gemma 4 12B romanized real sentences "the way Cambodians type on Messenger", with 4 style prompts. The Khmer side stays the original sentence. | |
| **Test separation.** | |
| - Every sentence of every evaluation set, including Samdei Type Bench, is removed from the corpus and banned in the generator. | |
| - No Samdei Type Bench message equals a training target. The only shared strings are common single words. | |
| **Reproducibility.** These files are the exact pairs the released models were trained on, so training can be repeated without regenerating them. The data generator and the training code are not public; the released code ([github.com/Samdei-Labs/samdei-type](https://github.com/Samdei-Labs/samdei-type)) is for inference only. | |
| ## Limitations | |
| - **Synthetic inputs.** No row is real typing; romanizations follow our four personas and one LLM. | |
| - **Pieces, not messages.** Inputs are short (median 26 characters, at most 132). | |
| - **Web noise.** About 1.5% of HPLT lines are misordered text extracted from PDFs. | |
| - **Personal data.** Web text may contain names, phone numbers or other personal data of the original authors. Lines with URLs or e-mail-like strings were dropped, but no other personal-data filter was run. To have a row removed, open a discussion in the [Community tab](https://huggingface.co/datasets/samdei/samdei-type-pretraining-data/discussions). | |
| - **LLM text.** The Gemma-written chat sentences and romanizations can contain mistakes. | |
| ## License and attribution | |
| **CC BY-SA 4.0** ([license text](https://creativecommons.org/licenses/by-sa/4.0/)). The rows contain pieces of the source sentences, so the source licenses travel with them: Khmer Wikipedia's share-alike terms set the license of the whole set, and rows from the other sources also stay under their own terms below. | |
| | Source | Side | License | | |
| |---|---|---| | |
| | HPLT v2 (`HPLT/HPLT2.0_cleaned`, khm_Khmr) | Khmer | CC0 packaging; Common Crawl terms of use apply to the text | | |
| | FineWeb-2 (`HuggingFaceFW/fineweb-2`, khm_Khmr) | Khmer | ODC-By 1.0; Common Crawl terms of use apply to the text | | |
| | Khmer Wikipedia (`wikimedia/wikipedia`, 20231101.km) | Khmer | CC BY-SA 4.0 (share-alike) | | |
| | Asian Language Treebank (ALT), Khmer side | Khmer | CC BY 4.0 | | |
| | LLM-written chat sentences (ours, Gemma 4) | Khmer | CC BY 4.0 (Gemma 4: Apache-2.0) | | |
| | Gemma 4 romanizations (ours) | Latin | Apache-2.0 model output; Khmer side as its source | | |
| | Google Khmer pronunciation lexicon | Latin (generator) | CC BY 4.0 | | |
| | Chat lexicon, place names, English chat words (ours) | Latin (generator) | CC0 | | |
| Rows whose `source` is `wikipedia` or `gemrom_wikipedia` carry Wikipedia's share-alike terms. They are about 19% of `train_web_s1`, `train_web_s2` and `train_v2_s3`. | |
| **Changes to the source text.** The sentences were normalized with `khnormal`, filtered and deduplicated; the generator cut them into pieces of 1–12 words. Each Khmer target is paired with a romanization in the `latin` column (see "How the pairs were made"). | |
| ## Citation | |
| ```bibtex | |
| @misc{samdei-type-2026, | |
| title = {Samdei Type: Separating What Is Written from How It Is Spelled for Romanized Khmer Transliteration}, | |
| author = {Mengsay Loem}, | |
| year = {2026}, | |
| note = {Technical report}, | |
| url = {https://github.com/Samdei-Labs/samdei-type/blob/main/paper/samdei-type.pdf} | |
| } | |
| ``` | |
| Contact: open a discussion in this repository's [Community tab](https://huggingface.co/datasets/samdei/samdei-type-pretraining-data/discussions). | |