vminibert-108M

vminibert-108M is a 108.6M-parameter Transformer encoder trained from scratch as a Masked Language Model on a blended 4-domain corpus (WikiText-103, BookCorpusOpen, FineWeb, and C4) using dynamic Whole-Word Masking (WWM).

Rather than a stock BERT reimplementation, it uses a modern LLaMA-style encoder recipe:

• RoPE (Rotary Position Embeddings) instead of learned absolute positions
• RMSNorm instead of LayerNorm
• GeGLU feed-forward blocks instead of a plain GELU MLP

Model Details

Parameters	108,615,984
Hidden size	768
Attention heads	12
Encoder layers	12
Intermediate size (GeGLU)	2048
Vocabulary size	30,000 (WordPiece)
Max sequence length	32
Positional encoding	RoPE
Normalization	RMSNorm
Activation	GeGLU
Weight tying	Output projection tied to input embeddings

Training Data

Trained on ~250k samples blended from four domains:

Source	Domain
WikiText-103	Encyclopedic
BookCorpusOpen	Long-form narrative
FineWeb	General web text
C4	Cleaned web crawl

Masking used dynamic Whole-Word Masking at a 15% rate.

Evaluation

Evaluated on official unseen test splits with standard 15% WWM:

Benchmark	Split	Masked Tokens	Loss	Perplexity	Top-1 Acc	Top-3 Acc	Top-5 Acc	MRR@10
Salesforce/wikitext-2-raw-v1	test	8,552	4.0323	56.39	43.15%	53.86%	57.96%	—
Salesforce/wikitext-103-raw-v1	test	8,515	4.0378	56.70	42.82%	53.55%	58.06%	49.29%

MRR@10 for wikitext-2 is not yet computed — fill in or remove the column before publishing.

Usage

from transformers import AutoTokenizer, AutoModelForMaskedLM, pipeline

model_id = "vprojectx/VMiniBert"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForMaskedLM.from_pretrained(model_id, trust_remote_code=True)

fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer)
results = fill_mask("The capital of France is [MASK].")

for res in results:
    print(f"{res['token_str']:<12} {res['score']:.4f}")

Limitations & Intended Use

• Context window: pretrained and evaluated at 32-token sequences only; behavior beyond that is untested. This is the single biggest caveat for anyone evaluating this model — flag it up front, not just in a bullet list.
• Corpus size: ~250k samples is modest for a from-scratch LM; expect it to need task-specific fine-tuning for downstream use rather than strong zero-shot performance.
• Intended use: a lightweight, fast MLM encoder for fine-tuning (classification, NER, masked-token extraction) and as a reference implementation of a modern (RoPE/RMSNorm/GeGLU) BERT-style architecture trained end-to-end from scratch.
• Not intended for: long-document tasks, production deployment without fine-tuning, or use cases requiring calibrated confidence at scale (no calibration study was done).

Citation

If you use this model, please cite it as:

@misc{vminibert2026,
  author = {vprojectx},
  title = {vminibert-108M: A From-Scratch BERT-Style Encoder with RoPE, RMSNorm, and GeGLU},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/vprojectx/VMiniBert}
}