Ego-153M-Modern

Ego-153M-Modern is a 153M-parameter decoder-only Transformer trained from scratch as a hands-on exploration of modern LLM architecture, optimization, and deployment workflows.

This project is intentionally positioned between research and engineering:

• Research-inspired architectural choices
• Production-grade training, checkpointing, and export
• Clean Hugging Face integration for reproducibility and inspection

The goal is deep systems understanding, not state-of-the-art benchmarks.

---

🔍 Motivation

After working extensively with API-based LLMs and RAG systems, I wanted to go deeper and understand:

• How architectural decisions (attention variants, normalization, activations) affect training dynamics
• How modern optimizers behave at scale
• How to correctly export and deploy custom Transformer architectures without forcing them into existing templates

Ego-153M-Modern is the result of that exploration.

---

🧠 Model Architecture

• Type: Decoder-only Transformer
• Total Parameters: ~153M
  (≈124M core Transformer + untied LM head + residual gating)
• Context Length: 1024 tokens
• Vocabulary: GPT-2 BPE (tiktoken compatible)

Key Architectural Choices

• Grouped Query Attention (GQA)
  Improves memory efficiency while retaining attention quality

• RoPE (Rotary Positional Embeddings)
  Precomputed and applied at attention time

• RMSNorm (no LayerNorm)
  Lower overhead and more stable scaling at depth

• ReLU² Activation
  Simple, fast nonlinearity explored in recent scaling work

• Untied Embeddings & LM Head
  Chosen deliberately for flexibility and experimentation

• Residual Gating Parameters
  Learnable residual scaling and input mixing per layer

• Logit Softcapping
  Stabilizes logits during training and early inference

---

⚙️ Training Setup

• Hardware: NVIDIA H200 GPU
• Precision: bfloat16 (Tensor Core optimized)
• Framework: PyTorch
• Compilation: torch.compile
• Context Length: 1024
• Effective Tokens / Step: ~262K

Optimizers

• Muon (Polar Express variant) for large linear layers
• Fused AdamW for embeddings, output head, and scalar parameters

This hybrid optimizer setup was chosen to explore convergence behavior in large matrix-heavy models.

---

📚 Dataset

• Source: FineWeb-Edu (Karpathy curated subset)
• Usage: Educational + experimental pretraining
• Goal: Architectural and systems learning, not benchmark dominance

---

📦 Model Format & Deployment

• Weights: safetensors
• Checkpoint Validation: CPU strict load + forward pass verification
• Hugging Face Integration: Custom architecture with manual instantiation
• Tokenizer: GPT-2 BPE (tiktoken compatible)

This repository intentionally avoids forcing the model into a GPT-2 or LLaMA template to preserve architectural fidelity.

---

⚠️ Limitations

• Not instruction-tuned
• No RLHF or preference alignment
• Limited world knowledge compared to frontier models
• Intended for experimentation, inspection, and extension

---

🔮 Future Work

• Instruction tuning and alignment
• Exploration of MoE / multi-latent attention variants
• KV-cache optimizations for inference
• Evaluation on structured reasoning benchmarks
• Integration with retrieval-augmented systems

---

👤 Author

Built by Ramesh Rathod as part of a broader journey across:

• LLM training from scratch
• Production-grade RAG systems
• Model export, evaluation, and deployment

Happy to discuss architecture, training trade-offs, and system design.

---

📎 License

For research and educational use.