pagantibet/Tibetan-KenLM-ACTibtrainingdata

Tibetan Normalisation - KenLM ACTib Training Data

A large-scale corpus of Standard Classical Tibetan text prepared specifically for training character-level KenLM n-gram language models for use in the PaganTibet normalisation pipeline. The dataset contains approximately 17.7 million lines of cleaned, line-split ACTib text in two versions: non-tokenised and tokenised (via a customised version of the Botok Tibetan tokeniser). These two files are the direct training corpora for pagantibet/5gram-kenLM_char and pagantibet/5gram-kenLM_char-tok respectively.

The KenLM models trained on this corpus serve as re-rankers for beam search candidates generated by the PaganTibet sequence-to-sequence normalisation models (tokenised S2S model and non-tokenised S2S model), selecting the hypothesis most consistent with fluent Standard Classical Tibetan.

This dataset is part of the PaganTibet project and accompanies the paper:

Meelen, M. & Griffiths, R.M. (2026) 'Historical Tibetan Normalisation: rule-based vs neural & n-gram LM methods for extremely low-resource languages' in Proceedings of the AI4CHIEF conference, Springer.

Please cite the paper and the code repository when using this dataset.

---

Dataset Overview

The dataset contains two plain text files, each representing the same underlying ACTib corpus (Annotated Corpus of Tibetan; >180 million words of Standard Classical Tibetan), available from Zenodo (Meelen & Roux 2020). Each file is prepared in a different form:

File	Size	Description
unsegACTib_lines_kenlm_20260103_220309.txt	2.16 GB	Non-tokenised ACTib lines — used to train pagantibet/5gram-kenLM_char
unsegACTib-tok_lines_kenlm_20260103_220119.txt	2.32 GB	Tokenised ACTib lines — used to train pagantibet/5gram-kenLM_char-tok

Note: the Hugging Face Dataset Viewer displays the dataset as a single train split — this is a technical default.

Both files are UTF-8 encoded plain text with one line per sequence. In the tokenised file, tokens are separated by spaces as produced by the customised Botok tokeniser (see Data_Preparation).

The ACTib does not contain natural linebreaks and includes non-Tibetan material (e.g. page numbers, section markers) that is unsuitable for language model training. Before use, the corpus was cleaned and split into artificial manuscript-length lines of varying length using the createTiblines.py script (see Data_Preparation). This produces lines of realistic, varied lengths that match the sequence units encountered during normalisation at inference time.

---

Usage

The training notebook KenLM_trainforNormalisation.ipynb walks through the full training process and can be run on Google Colab. The key steps are as follows:

Step 1 — Install KenLM

sudo apt-get install -y build-essential cmake libboost-all-dev \
    libboost-program-options-dev libboost-system-dev \
    libboost-thread-dev libboost-test-dev libeigen3-dev zlib1g-dev
git clone https://github.com/kpu/kenlm.git
cd kenlm && mkdir build && cd build
cmake .. && make -j 4

Step 2 — Convert to Space-Separated Characters

KenLM requires space-separated tokens. For character-level modelling, each character must be separated by a space. Apply this to whichever file you are using:

# For the non-tokenised file
with open("unsegACTib_lines_kenlm_20260103_220309.txt") as f_in, \
     open("actib_char.txt", "w") as f_out:
    for line in f_in:
        f_out.write(" ".join(line.strip()) + "\n")

# For the tokenised file
with open("unsegACTib-tok_lines_kenlm_20260103_220119.txt") as f_in, \
     open("actib_char_tok.txt", "w") as f_out:
    for line in f_in:
        f_out.write(" ".join(line.strip()) + "\n")

Step 3 — Train the 5-gram KenLM

Using the parameters from Meelen & Griffiths (2026):

# Non-tokenised model
./kenlm/build/bin/lmplz \
    -o 5 \
    --prune 0 1 1 2 2 \
    --discount_fallback \
    < actib_char.txt \
    > model_5gram_char.arpa

# Tokenised model
./kenlm/build/bin/lmplz \
    -o 5 \
    --prune 0 1 1 2 2 \
    --discount_fallback \
    < actib_char_tok.txt \
    > model_5gram_char_tok.arpa

Parameter notes:

• -o 5 — 5-gram order
• --prune 0 1 1 2 2 — pruning thresholds per n-gram order (unigrams kept, higher-order singletons pruned)
• --discount_fallback — modified Kneser-Ney discounting with fallback

Training completes in under 5 minutes on Google Colab. Full parameter settings are reported in the Appendix of Meelen & Griffiths (2026).

Step 4 — (Optional) Convert to Binary

For faster loading with the compiled KenLM Python backend:

./kenlm/build/bin/build_binary model_5gram_char.arpa model_5gram_char.bin

Note: The pure Python ARPA backend used in the inference scripts requires .arpa format. The .bin format is only compatible with the compiled kenlm Python package.

---

Relationship to Other Resources

This corpus sits at the base of the KenLM training pipeline. The diagram below shows how it connects to the rest of the PaganTibet normalisation stack:

ACTib corpus (Zenodo)
       │
       ▼
createTiblines.py  ──→  unsegACTib_lines_kenlm_*.txt       (this dataset, non-tokenised file)
       │
       ▼
botokenise_src-tgt.py  ──→  unsegACTib-tok_lines_kenlm_*.txt  (this dataset, tokenised file)
       │
       ├──→  lmplz (KenLM)  ──→  5gram-kenLM_char        (non-tokenised KenLM model)
       └──→  lmplz (KenLM)  ──→  5gram-kenLM_char-tok    (tokenised KenLM model)
                                        │
                                        ▼
                              Used as beam search re-ranker
                              in normalisationS2S inference

---

Related Models and Resources

Resource	Link
Non-tokenised KenLM model (trained on this data)	pagantibet/5gram-kenLM_char
Tokenised KenLM model (trained on this data)	pagantibet/5gram-kenLM_char-tok
Non-tokenised Seq2Seq model	pagantibet/normalisationS2S-nontokenised
Tokenised Seq2Seq model	pagantibet/normalisationS2S-tokenised
S2S training dataset	pagantibet/normalisation-S2S-training
KenLM training notebook	KenLM_trainforNormalisation.ipynb
Data preparation scripts & ReadMes	github.com/pagantibet/normalisation/Data_Preparation
Inference scripts & ReadMe	github.com/pagantibet/normalisation/Inference
ACTib corpus	Zenodo (Meelen & Roux 2020)
PaganTibet project	pagantibet.com

---

License

This dataset is released under CC BY-NC-SA 4.0. It may be used freely for non-commercial research and educational purposes, with attribution and under the same licence terms.

---

Funding

This work was partially funded by the European Union (ERC, Pagan Tibet, grant no. 101097364). Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency.