kd13/RoPERT-MLM-mini

BERTmini — Custom BERT with RoPE & Pre-LN Trained from Scratch

A compact BERT-style masked language model trained entirely from scratch on BookCorpus. The architecture replaces the canonical absolute positional embeddings with Rotary Position Embeddings (RoPE) and adopts a Pre-Layer Normalization (Pre-LN) residual layout, both of which have become standard practice in modern transformer training.

---

Architecture Design Choices

RoPE instead of learned absolute position embeddings. Rotary embeddings encode positional information directly into the query–key dot product, enabling length generalisation beyond the training window and eliminating a separate learnable parameter table.

Pre-LN residual stream. Layer normalisation is applied to the input of each sub-layer rather than the output. This stabilises gradient flow during early training and generally makes the loss curve smoother, at the cost of requiring an explicit final encoder normalisation before the prediction head.

Embedding tying. The MLM decoder projection matrix shares weights with the token embedding table, which reduces parameter count and typically improves token prediction quality.

---

Training Details

Dataset

Split	Source	Packing
Train	BookCorpus	Fixed-length packed sequences (128 tokens)
Eval	BookCorpus (held-out)	Same packing strategy

Sequences were packed (not padded) to maximise token utilisation per batch, following the approach used in the original BERT paper.

Pre-training Objective

Masked Language Modelling (MLM) with a masking probability of 30 %. The standard 80/10/10 mask/replace/keep strategy is applied by DataCollatorForLanguageModeling.

Evalution Details

Perplexity : 12.59

---

Usage

import torch
from transformers import AutoTokenizer, AutoModelForMaskedLM

model_name = "kd13/RoPERT-MLM-mini"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained(model_name, trust_remote_code=True)

text = "i don't have much [MASK]."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

mask_token_index = torch.where(inputs["input_ids"] == tokenizer.mask_token_id)[1]
mask_token_logits = logits[0, mask_token_index, :]
top_5_tokens = torch.topk(mask_token_logits, 5, dim=1).indices[0].tolist()

for token in top_5_tokens:
    print(f">>> {text.replace('[MASK]', tokenizer.decode([token]))}")

---

Limitations

• Domain coverage. The model was pre-trained exclusively on BookCorpus (narrative English text). Performance on technical, scientific, or conversational text may be weaker than models trained on broader corpora such as Wikipedia + BookCorpus (original BERT) or C4 (RoBERTa-style).

• Sequence length. The RoPE cache and packing strategy are fixed at 128 tokens. While RoPE theoretically supports length extrapolation, the model has not been tested beyond this limit.

• Scale. At ~20 M parameters, this model is best suited for learning experiments, fine-tuning research, or resource-constrained deployment scenarios. It is not designed to compete with bert-base-uncased (110 M) or larger checkpoints on downstream benchmarks.

• No fine-tuning benchmarks. GLUE/SuperGLUE evaluation has not been performed. Perplexity on the held-out BookCorpus split (PPL ≈ 12.5) is the only reported metric at this stage.

---