Panhapich/khmer-llada-shards

Khmer-LLaDA pre-tokenized shards

Packed uint16 training shards for Khmer-LLaDA-Small (small_a / small_b) — pre-tokenized so a Kaggle (or any) training run can download and start training directly, with no corpus download, no khmer-nltk segmentation pass, and nothing that can drop a connection mid-tokenize.

What's in here

files	source	tokens
train_0000.npy .. train_0014.npy, val_0000.npy	original corpus (Panhapich/khmer-text-corpus, all_text_segmented.txt)	146.1M
train_0015.npy .. train_0040.npy	extra corpus: nphearum/khmer-raw-text-3M-v2 + wikimedia/wikipedia (20231101.km)	261.7M
total		407.8M

Both small_a and small_b train on this exact same shard set — architecture (configs/small_a.json vs configs/small_b.json) is the only intended difference between the two runs, so results are comparable head to head.

Format

Each .npy file is a packed uint16 array of shape (n_sequences, 512) — tokenized with Panhapich/khmer-sp-8k (vocab size 8000; <PAD>=0, <UNK>=1, <BOS>=2, <EOS>=3, <MASK>=4). Rows are <EOS>-delimited packed token streams truncated to seq_len=512, matching what scripts/pretokenize.py produces. A random ~1% of TRAIN rows are truncated to a random length in [1, 512) and <PAD>-filled, per the LLaDA paper's variable-length trick (§2.2) — see apply_variable_length in that script for the exact rationale.

val_0000.npy is held out from the END of the original (shuffled) corpus only — the extra corpus contributes no validation data, so val_nll_bound stays comparable across the whole run history (before and after the extra corpus was added).

How it was built

1. scripts/pretokenize.py --in data/raw/all_text_segmented.txt --seq-len 512 (original corpus -> train_0000..0014.npy + val_0000.npy)
2. scripts/extract_extra_corpus.py (nphearum + Khmer Wikipedia -> combined_raw.txt)
3. scripts/pretokenize_extra_parallel.py --workers 72 (khmer-nltk word-segmentation + tokenize, appended as train_0015..0040.npy)

Usage

from huggingface_hub import snapshot_download
snapshot_download("Panhapich/khmer-llada-shards", repo_type="dataset",
                   local_dir="data/shards", allow_patterns=["*.npy"])

Then point training/train.py at data/shards/train_*.npy / data/shards/val_*.npy as usual (see configs/train_t4.yaml / configs/train_t4_b.yaml).