DipeshChaudhary/nepali-gector-style-token-level-tag

Nepali GEC (gector style) Token Tagging Dataset

This is a processed version of the sumitaryal/nepali_grammatical_error_correction dataset, designed for training GEC-ToR-style sequence tagging models.

This dataset has been processed with a robust, multi-pass, content-aware alignment algorithm to generate high-fidelity correction tags, including complex and adjacent SWAP operations in token level usng iriiS-RESEARCH/RoBERTa_Nepali_125M.

Total Examples: 16,260,992

• Training: 13,008,711
• Validation: 2,439,231
• Calibration: 813,050

Key Features

• Token-Level Tags: Word-level corrections are mapped to subword (token) labels.
• Correct Sentences Included: The dataset contains both incorrect and correct sentences (tagged with $KEEP) for model stability.
• Stratified Splits: Splits are stratified by the dominant error tag to ensure balanced evaluation.

Tag Vocabulary (10 Tags)

This dataset uses an enhanced 10-tag system:

1. $KEEP: Token is correct. -> label -> 0
2. $DELETE: Token should be deleted. -> label -> 1
3. $REPLACE: Token should be replaced (e.g., by a Transformer's MLM head). -> label -> 2
4. $APPEND: A new token should be inserted after this token. -> label -> 3
5. $SWAP_NEXT: Token is part of a swap (first word). -> label -> 4
6. $SWAP_PREV: Token is part of a swap (second word). -> label -> 5
7. $MERGE_NEXT: This token should be merged with the next token. -> label -> 6
8. $MERGE_PREV: This token should be merged with the previous token. -> label -> 7
9. $SPLIT: This token should be split into multiple tokens. -> label -> 8
10. $UNKNOWN: Fallback tag (should not be present). -> label -> 9

Usage

from datasets import load_dataset
import json
from huggingface_hub import hf_hub_download

REPO_ID = "DipeshChaudhary/nepali-gector-style-token-level-tag"

# Load dataset
full_dataset = load_dataset(
    "REPO_ID",
    split=["train", "validation", "calibration"]
)

# Load vocabulary
vocab_file = hf_hub_download(
    repo_id=REPO_ID,
    filename="gec_vocabulary.json",
    repo_type="dataset"
)

with open(vocab_file, 'r', encoding='utf-8') as f:
    vocabulary = json.load(f)

print(f"Splits: {list(full_dataset.keys())}")
print(f"Tags: {vocabulary['tag_to_id']}")



# full_dataset will be a list containing the three loaded Dataset objects
train_ds = full_dataset[0]
valid_ds = full_dataset[1]
calib_ds = full_dataset[2]

print(train_ds)
print(valid_ds)
print(calib_ds)