BERT-PRETRAINED-EDU

Overview

This is a BERT-style masked language model pretrained from scratch using streaming data from FineWeb-Edu.

Architecture

• Hidden size: 384
• Layers: 6
• Heads: 6
• Sequence length: 128
• Objective: Masked Language Modeling (MLM)

Training

• Dataset: HuggingFaceFW/fineweb-edu (streaming)
• Steps: ~20,000
• GPUs: Dual GPU (DDP)
• Mixed Precision Training

Intended Use

• Fine-tuning for:
  • Sentiment classification
  • Document classification
  • Retrieval / RAG encoder
  • NLP research

Limitations

• Not instruction-tuned
• Not chat-optimized

Use Case

1. Masked Language Modeling

import torch
from transformers import AutoTokenizer

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    "your_hf_username/bert-edu-pretrained-384d"
)

# Load model weights (custom architecture)
from model import BERT, BERTConfig  # your model definition

config = BERTConfig(
    vocab_size=30522,
    dim=384,
    n_layers=6,
    n_heads=6,
    seq_len=128
)

model = BERT(config)
model.load_state_dict(
    torch.load("pytorch_model.bin", map_location="cpu")
)
model.eval()

# Input with [MASK]
text = "Machine learning is [MASK]."
inputs = tokenizer(text, return_tensors="pt")

# Forward pass
with torch.no_grad():
    logits = model(
        inputs["input_ids"],
        torch.zeros_like(inputs["input_ids"])
    )

mask_index = (inputs["input_ids"] == tokenizer.mask_token_id).nonzero()[0, 1]
pred_id = logits[0, mask_index].argmax(dim=-1)

print("Prediction:", tokenizer.decode(pred_id))

Finetuning for Sentiment Classification

import torch.nn as nn

class BertForSentiment(nn.Module):
    def __init__(self, bert, hidden_size, num_labels=2):
        super().__init__()
        self.bert = bert
        self.classifier = nn.Linear(hidden_size, num_labels)

    def forward(self, input_ids, attention_mask=None):
        seg = torch.zeros_like(input_ids)
        hidden_states = self.bert(input_ids, seg)
        cls_token = hidden_states[:, 0, :]
        return self.classifier(cls_token)

# Step-2 :
bert = BERT(config)
bert.load_state_dict(
    torch.load("pytorch_model.bin", map_location="cpu")
)

model = BertForSentiment(bert, hidden_size=384)

# train onto sentiemental data(imdb)
from datasets import load_dataset

dataset = load_dataset("imdb")

# tokenize → train → evaluate