RexBERT-base

RexBERT-base

TL;DR: An encoder-only transformer (ModernBERT-style) for e-commerce applications, trained in three phases—Pre-training, Context Extension, and Decay—to power product search, attribute extraction, classification, and embeddings use cases. The model has been trained on 2.3T+ tokens along with 350B+ e-commerce-specific tokens

---

Table of Contents

• Quick Start
• Intended Uses & Limitations
• Model Description
• Training Recipe
• Data Overview
• Evaluation
• Usage Examples
  • Masked language modeling
  • Embeddings / feature extraction
  • Text classification fine-tune
• Model Architecture & Compatibility
• Efficiency & Deployment Tips
• Responsible & Safe Use
• License
• Maintainers & Contact
• Citation

---

Quick Start

import torch
from transformers import AutoTokenizer, AutoModel, AutoModelForMaskedLM, pipeline

MODEL_ID = "thebajajra/RexBERT-base"

# Tokenizer
tok = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)

# 1) Fill-Mask (if MLM head is present)
mlm = pipeline("fill-mask", model=MODEL_ID, tokenizer=tok)
print(mlm("These running shoes are great for [MASK] training."))

# 2) Feature extraction (CLS or mean-pooled embeddings)
enc = AutoModel.from_pretrained(MODEL_ID)
inputs = tok(["wireless mouse", "ergonomic mouse pad"], padding=True, truncation=True, return_tensors="pt")
with torch.no_grad():
    out = enc(**inputs, output_hidden_states=True)
# Mean-pool last hidden state for sentence embeddings
emb = (out.last_hidden_state * inputs.attention_mask.unsqueeze(-1)).sum(dim=1) / inputs.attention_mask.sum(dim=1, keepdim=True)

---

Intended Uses & Limitations

Use cases

• Product & query retrieval/semantic search (titles, descriptions, attributes)
• Attribute extraction / slot filling (brand, color, size, material)
• Classification (category assignment, unsafe/regulated item filtering, review sentiment)
• Reranking and query understanding (spelling/ASR normalization, acronym expansion)

Out of scope

• Long-form generation (use a decoder/seq-to-seq LM instead)
• High-stakes decisions without human review (pricing, compliance, safety flags)

Target users

• Search/recs engineers, e-commerce data teams, ML researchers working on domain-specific encoders

---

Model Description

RexBERT-base is an encoder-only, 150M parameter transformer trained with a masked-language-modeling objective and optimized for e-commerce related text. The three-phase training curriculum improves general language understanding, extends context handling, and then specializes on a very large corpus of commerce data to capture domain-specific terminology and entity distributions.

---

Training Recipe

RexBERT-base was trained in three phases:

1. Pre-training
   General-purpose MLM pre-training on diverse English text for robust linguistic representations.

2. Context Extension
   Continued training with increased max sequence length to better handle long product pages, concatenated attribute blocks, multi-turn queries, and facet strings. This preserves prior capabilities while expanding context handling.

3. Decay on 350B+ e-commerce tokens
   Final specialization stage on 350B+ domain-specific tokens (product catalogs, queries, reviews, taxonomy/attributes). Learning rate and sampling weights are annealed (decayed) to consolidate domain knowledge and stabilize performance on commerce tasks.

Training details (fill in):

• Optimizer / LR schedule: TODO
• Effective batch size / steps per phase: TODO
• Context lengths per phase (e.g., 512 → 1k/2k): TODO
• Tokenizer/vocab: TODO
• Hardware & wall-clock: TODO
• Checkpoint tags: TODO (e.g., pretrain, ext, decay)

---

Data Overview

• Dataset: Ecom-niverse
• Domain mix:

We identified 9 E-commerce overlapping domains which have significant amount of relevant tokens but required filteration. Below is the domain list and their filtered size

Additionally, there are 6 more domains which had almost complete overlap and were picked directly out of FineFineWeb.

By focusing on these domains, we narrow the search space to parts of the web data where shopping-related text is likely to appear. However, even within a chosen domain, not every item is actually about buying or selling, many may be informational articles, news, or unrelated discussions. Thus, a more fine-grained filtering within each domain is required to extract only the e-commerce-specific lines. We accomplish this by training lightweight classifiers per domain to distinguish e-commerce context vs. non-e-commerce content.

---

Evaluation

Token Classification

With 2–3x fewer parameters, RexBERT surpasses the performance of the ModernBERT series.

Semantic Similarity

RexBERT models outperform all the models in their parameter/size category.

---

Usage Examples

1) Masked language modeling

from transformers import AutoModelForMaskedLM, AutoTokenizer, pipeline

m = AutoModelForMaskedLM.from_pretrained("thebajajra/RexBERT-base")
t = AutoTokenizer.from_pretrained("thebajajra/RexBERT-base")
fill = pipeline("fill-mask", model=m, tokenizer=t)

fill("Best [MASK] headphones under $100.")

2) Embeddings / feature extraction

import torch
from transformers import AutoTokenizer, AutoModel

tok = AutoTokenizer.from_pretrained("thebajajra/RexBERT-base")
enc = AutoModel.from_pretrained("thebajajra/RexBERT-base")

texts = ["nike air zoom pegasus 40", "running shoes pegasus zoom nike"]
batch = tok(texts, padding=True, truncation=True, return_tensors="pt")

with torch.no_grad():
    out = enc(**batch)
# Mean-pool last hidden state
attn = batch["attention_mask"].unsqueeze(-1)
emb = (out.last_hidden_state * attn).sum(1) / attn.sum(1)
# Normalize for cosine similarity (recommended for retrieval)
emb = torch.nn.functional.normalize(emb, p=2, dim=1)

3) Text classification fine-tune

from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer

tok = AutoTokenizer.from_pretrained("thebajajra/RexBERT-base")
model = AutoModelForSequenceClassification.from_pretrained("thebajajra/RexBERT-base", num_labels=NUM_LABELS)

# Prepare your Dataset objects: train_ds, val_ds (text→label)
args = TrainingArguments(
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    learning_rate=3e-5,
    num_train_epochs=3,
    evaluation_strategy="steps",
    fp16=True,
    report_to="none",
    load_best_model_at_end=True,
)

trainer = Trainer(model=model, args=args, train_dataset=train_ds, eval_dataset=val_ds, tokenizer=tok)
trainer.train()

---

Model Architecture & Compatibility

• Architecture: Encoder-only, ModernBERT-style base model.
• Libraries: Works with 🤗 Transformers; supports fill-mask and feature-extraction pipelines.
• Context length: Increased during the Context Extension phase—ensure max_position_embeddings in config.json matches your desired max length.
• Files: config.json, tokenizer files, and (optionally) heads for MLM or classification.
• Export: Standard PyTorch weights; you can export ONNX / TorchScript for production if needed.

---

Responsible & Safe Use

• Biases: Commerce data can encode brand, price, and region biases; audit downstream classifiers/retrievers for disparate error rates across categories/regions.
• Sensitive content: Add filters for adult/regulated items; document moderation thresholds if you release classifiers.
• Privacy: Do not expose PII; ensure training data complies with terms and applicable laws.
• Misuse: This model is not a substitute for legal/compliance review for listings.

---

License

• License: apache-2.0.

---

Maintainers & Contact

• Authors: Rahul Bajaj, Anuj Garg

---