README-2.md

---
license: cc-by-nc-4.0
language:
  - en
tags:
  - babylm
  - babylm-2026
  - mixture-of-experts
  - msit
  - xpertgpt
  - swiglu
  - custom_code
  - safetensors
library_name: transformers
pipeline_tag: text-generation
---

# XpertGPT (SwiGLU & Sliding Window 64, 16, 8, 4)

XpertGPT is a sparse Mixture of Experts (MoE) language model designed for data-efficient pretraining under the **BabyLM 2026 challenge (Strict-Small 10M track)**. It leverages **Parallelized Multi-Scale Information Transmission (MSIT)** and **Expert Choice Routing** to maximize representational capacity within restricted token budgets.

This version implements **SwiGLU Feed-Forward Networks** across all global and parallel expert blocks, along with corrected LayerNorms, redundant residual removal, and **sliding window attention sizes** of `[64, 16, 8, 4]` tokens.

---

## 1. SwiGLU Feed-Forward Networks (FFN)

Instead of the standard Feed-Forward sequential structure (Linear -> GELU -> Linear), this model replaces FFN layers with the **SwiGLU (Swish Gated Linear Unit)** variant to improve model capacity and training stability:

$$\text{FFN}_{\text{SwiGLU}}(x) = \left(\text{Swish}(x W) \otimes x V\right) W_2$$

Where the Swish function is implemented using SiLU:
* **Gate Linear projections ($W$, $V$)**: Projects input dimension `dim` to `hidden_dim`.
* **Out Linear projection ($W_2$)**: Projects back to `dim`.
* To maintain parameter counts equivalent to standard `dim * 4` sequential FFNs, the hidden dimension is scaled to:
  $$\text{hidden\_dim} = \text{round\_to\_multiple\_of\_8}\left(\frac{8}{3} \times \text{dim}\right)$$

---

## 2. Expert Sliding Window Layout

The four parallel MoE experts are configured with distinct sliding window attention constraints:
* **Expert 1**: Window size `64` tokens
* **Expert 2**: Window size `16` tokens
* **Expert 3**: Window size `8` tokens
* **Expert 4**: Window size `4` tokens

---

## 3. Architectural Layout & Changes

This model implements:
1. **Removal of Redundant Residual (`res3`)**:
   * Removed redundant residual connection around the global dense block. Gated input is now simply $X_2 = X_1$.
2. **Introduction of Post-Block LayerNorm (`ln3`)**:
   * LayerNorm `ln3` is added after the SwiGLU addition inside every `MSITBranchBlock`.
   * Formulation: $X_{\text{out}} = \text{LayerNorm}(X^{(2)})$.
3. **Introduction of Post-MoE LayerNorm (`ln_post_moe`)**:
   * LayerNorm `ln_post_moe` is added after the Residual 4 MoE aggregation.
   * Formulation: $X_{\text{out}} = \text{LayerNorm}(X_2 + X_{3, \text{full}})$.

---

## 4. How to Load and Use Checkpoints (Bypass Retraining)

### A. Loading the Final Model (`main` branch)
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "SRJ5035/sw_glu_sw_64_16_8_4_xpert_gpt",
    revision="main",
    trust_remote_code=True
).eval()

tokenizer = AutoTokenizer.from_pretrained(
    "SRJ5035/sw_glu_sw_64_16_8_4_xpert_gpt",
    revision="main"
)
```

### B. Loading an Intermediate Milestone (e.g. `chck_5M`)
```python
model_5m = AutoModelForCausalLM.from_pretrained(
    "SRJ5035/sw_glu_sw_64_16_8_4_xpert_gpt",
    revision="chck_5M",
    trust_remote_code=True
).eval()
```