Darja-GPT-50M-Padded-Data

Darja-GPT-50M-Padded-Data is the official pre-tokenized, chunked, and domain-conditioned dataset used to pre-train Darja-GPT-50M.

It processes the raw conversational transcripts from the Algerian Darja Corpus (conditioned subset) into fixed 1,024-token sequences with custom Byte-Level BPE tokens, explicit attention_mask tensors, target labels, and per-sequence domain_ids for multi-domain adapter training.

---

---

Dataset Overview & Motivation

Training a causal language model from scratch on dialectal transcripts poses unique challenges:

1. The Truncation Waste Problem: Naively truncating long podcast or vlog transcripts at 1,024 tokens discards over 80% of conversational content.
2. The Document Bleed Problem: Standard group_texts concatenates multiple unrelated documents into one long stream. In domain-conditioned training, this causes cooking recipes to bleed into football match discussions within the same context window.
3. Dynamic Padding Inefficiency: Dynamic batch padding on Kaggle GPU instances introduces variable memory usage, risk of CUDA OOMs, and recompilation overhead.

This dataset solves all three problems by sequentially chunking each document individually, padding non-full trailing chunks to a static 1,024 length with -100 label masking, and attaching metadata domain_ids directly to the tensor structure.

---

Data Pipeline & Preprocessing Engineering

The dataset was constructed from 11,151 documents of the conditioned subset of touati-kamel/algerian-darja-corpus:

1. Reproducible Split: Split into 97% training (10,816 documents) and 3% held-out evaluation (335 documents) using fixed seed 42.
2. Domain Merge Rule: <comedy> and <lifestyle_vlog> were consolidated into <general> to concentrate colloquial expressions and eliminate memorization risks on small document counts.
3. Tokenization: Tokenized with touati-kamel/darja-tokenizer-50m (Vocab size 24,000, Byte-Level BPE).
4. Sequential Chunking:
   • Each encoded document is sliced into consecutive chunks of 1,024 tokens: [0:1024], [1024:2048], etc.
   • Any short final chunk (<5 tokens) is discarded.
   • Chunks shorter than 1,024 tokens are padded with tokenizer.pad_token_id (0).
   • Yield: Expanded the training corpus from 10,816 raw documents into 52,038 dense 1,024-token training examples (~4.8× usable data yield).

---

Dataset Schema & Features

Each example in train and eval contains the following pre-computed fields:

Feature	Type	Shape	Description
input_ids	Sequence(int32)	[1024]	Token IDs encoded with darja-tokenizer-50m, right-padded with 0
attention_mask	Sequence(int8)	[1024]	Binary mask: 1 for valid conversational tokens, 0 for padding tokens
domain_ids	int64	Scalar	Integer domain index [0, 5] specifying which LoRA adapter to activate
labels	Sequence(int64)	[1024]	Causal LM target tokens. Matches input_ids for valid tokens, set to -100 for pad tokens

---

Domain Distribution & ID Mapping

The scalar domain_ids feature maps to the 6 primary domain adapters of Darja-GPT-50M:

Domain ID	Tag	Description	Representative Source Content
0	<general>	General conversation, comedy, vlogs	DZjoker, Anes Tina, Mourad Oudia, everyday life vlogs
1	<cooking>	Traditional & modern culinary arts	Amina Cuisine, Chef Moha, recipes, kitchen terms
2	<tech>	Electronics, gadgets, technology	Hardware benchmarks, mobile reviews, e-commerce
3	<sports>	Sports commentary, Algerian football	Dzair Foot, Dz Sport tv, national team, Mahrez, leagues
4	<podcast>	Business, marketing, self-help	Entrepreneurship podcasts, career mindset, freelancing
5	<story>	Narrative folklore & social tales	Storytelling, family chronicles, traditional stories

(Note: During model training, an additional virtual domain 6: <NULL> was dynamically introduced with 10% dropout to enable Classifier-Free Guidance).

---

Text-Tag Dropout Regularization

To ensure that the domain adapters learn genuine semantic and syntactic domain features rather than simply memorizing the prompt's prefix tag:

• 20% of training examples had the literal text tag (e.g. <cooking> ) stripped from the beginning of the text during processing.
• Crucially, the domain_ids tensor remained set to the correct domain index (1).
• This forced the attention and MLP adapters to model the vocabulary, register, and grammar of the domain directly from the conversational content.

---

Quickstart & Usage Example

1. Installation

pip install datasets transformers torch

2. Loading the Dataset & PyTorch Formatting

from datasets import load_dataset
from transformers import AutoTokenizer

DATASET_ID = "touati-kamel/darja-gpt-50m-padded-data"
TOKENIZER_ID = "touati-kamel/darja-tokenizer-50m"

# 1. Load dataset splits
dataset = load_dataset(DATASET_ID)
print("Splits:", dataset)
print("Train examples:", len(dataset["train"]))
print("Eval examples:", len(dataset["eval"]))

# 2. Set format for PyTorch
cols = ["input_ids", "attention_mask", "domain_ids", "labels"]
dataset["train"].set_format(type="torch", columns=cols)
dataset["eval"].set_format(type="torch", columns=cols)

# 3. Inspect a sample batch
sample = dataset["train"][0]
print("\n--- Sample 0 ---")
print("Input IDs shape:", sample["input_ids"].shape)
print("Domain ID:", sample["domain_ids"].item())
print("Valid tokens:", sample["attention_mask"].sum().item())
print("Padding tokens:", (sample["labels"] == -100).sum().item())

# 4. Decode the content
tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_ID)
valid_ids = sample["input_ids"][:sample["attention_mask"].sum()]
print("\nDecoded text preview:")
print(tokenizer.decode(valid_ids, skip_special_tokens=False)[:300], "...")

3. Using Directly with PyTorch DataLoader

from torch.utils.data import DataLoader

train_loader = DataLoader(
    dataset["train"],
    batch_size=16,
    shuffle=True,
    num_workers=2,
    pin_memory=True
)

for batch in train_loader:
    # Directly feed into DarjaGPTForCausalLM
    # out = model(
    #     input_ids=batch["input_ids"],
    #     domain_ids=batch["domain_ids"],
    #     attention_mask=batch["attention_mask"],
    #     labels=batch["labels"]
    # )
    break

---

Citation & Contact

If you use Darja-GPT-50M-Padded-Data or the Algerian Darja Corpus, please cite:

@misc{touati2026darjapaddeddata,
  author       = {Kamel Touati},
  title        = {Darja-GPT-50M-Padded-Data: Pre-Tokenized and Chunked Domain-Conditioned Dataset for Algerian Darija},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/touati-kamel/darja-gpt-50m-padded-data}}
}

@misc{touati2026darjagpt50m,
  author       = {Kamel Touati},
  title        = {Darja-GPT-50M: Domain-Conditioned Causal Language Model for Algerian Darija with Classifier-Free Guidance},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/touati-kamel/darja-gpt-50m}}
}

Developer: Kamel Touati (Hugging Face Profile)
License: Apache 2.0