Long Context XML-Roberta-large

A long‑context variant of XLM‑RoBERTa Large extended to 4096 tokens and continued‑pretrained on CC100 with RoBERTa‑style sequence packing and dynamic masking (MLM).

• Context window: 4096 tokens (implemented as max_position_embeddings ≈ 4098, i.e., 4096 effective tokens + specials)
• Objective: Masked Language Modeling (MLM) with mlm_probability=0.15
• Data: CC100 (all languages) via 🤗 Datasets streaming, weigthed interleaved across languages
• Packing: concatenate documents and insert </s> separators. Fixed‑length blocks.
• Positional embeddings: extended from the base model using interpolation then full model finetuned.

Intended uses

• General multilingual text representation and MLM pre‑training tasks
• Backbone for multilingual classification, retrieval, and sequence tagging
• Starting point for continued task‑specific fine‑tuning that benefits from longer context (documents, code, QA context passages)

How to use:

from transformers import (
    pipeline,
    AutoTokenizer,
    AutoModelForMaskedLM,
)
ckpt = "pierre-tassel/longcontext-xlm-roberta-large-4096-tokens"
tokenizer = AutoTokenizer.from_pretrained(ckpt)
model = AutoModelForMaskedLM.from_pretrained(ckpt)

text = "Paris is the capital of <mask>."
fill = pipeline("fill-mask", model=ckpt, tokenizer=ckpt)
print(fill(text))

Model architecture

• Base: FacebookAI/xlm-roberta-large
• Tokenizer: XLM‑R SentencePiece
• Max positions: extended to ~4098 (effective 4096 tokens usable)

Data

• Source: cc100 (all available language configs)
• Access mode: streaming=True
• Mixing: per‑language shuffled streams, then weighted sampling (alpha=0.3)
• Preprocessing: whitespace trim
• Packing: concatenate tokenized docs; insert between docs; emit fixed‑length blocks Because of interleaving and packing, individual blocks and batches can contain multiple languages.

Objective & masking

• Data collator: DataCollatorForLanguageModeling
• MLM with mlm_probability=0.15 Optimizer & schedule
• Optimizer: adamw_torch_fused
• Learning rate: 1e-4
• Weight decay: 0.01
• Warmup ratio: 0.005
• Max grad norm: 1.0 Sequence‑length curriculum
• Start block size: 768
• Increment: +256 tokens every 1024 steps
• Target: 4096

Positional embedding extension

• Extends absolute positional embeddings from the base model to the new max positions
• Method: interpolation of non‑pad rows with per‑dimension mean/std matching of the tail
• Preserves pad index semantics and gradient settings