Wigip_v2_1.7B_ViT

WIGIP-1 v2

WIGIP-1 v2

⚠️ implementation & Training Scripts: The full source code, JAX training loops, and architecture definitions are available on my GitHub: 🔗 Click here to view the Training Scripts on GitHub

---

Stage 1 – Text Pre-Training (ViT-Style Transformer)

WIGIP-1 v2 is an experimental research model exploring Vision Transformer (ViT) style architectures for text modeling, implemented using JAX + Flax with Fully Sharded Data Parallelism (FSDP) via pjit.

This repository currently contains ONLY Phase 1 (Text Pre-Training).

---

⚠️ Training Status (IMPORTANT)

• ✅ Phase 1: Text-only pre-training

  • Character-level language modeling
  • Dataset: C4 (English)
  • Architecture: ViT-style transformer applied to reshaped text
  • ~57,000 training steps completed
  • Training performed using streaming data and FSDP

• ❌ Phase 2: Image training (NOT DONE)

  • No image data has been used
  • No multimodal or vision supervision yet
  • This phase is planned for future work

🚨 Model weights will be updated in the future once Phase 2 training is performed. Do NOT treat current checkpoints as a final or multimodal-capable model.

---

🧠 Model Overview (Phase 1)

• Text is tokenized at character level
• Tokens are reshaped into a 2D grid
• Grid is treated like an image and processed using:
  • Patch embedding via convolution
  • Multi-head self-attention
  • Feed-forward blocks
• Final output predicts the next character token

This phase is intended to test whether ViT-style inductive biases can learn meaningful structure from text alone.

---

⚙️ Technical Highlights

• JAX + Flax + Optax
• pjit with 2D mesh (data, model)
• Activation rematerialization (nn.remat)
• Gradient clipping
• Warmup + cosine learning rate schedule
• Streaming dataset (no full dataset in memory)

---

💾 Checkpointing

• Checkpoints are:
  • Automatically saved at time intervals
  • Compressed into .zip archives
  • Contain:
    • Model parameters (.pkl.gz)
    • Optimizer state
    • Training step metadata
• Training can be safely resumed from the latest zipped checkpoint

---

🔮 Future Work

• Phase 2: Image-based training
• Multimodal alignment (text + vision)
• Scaling beyond current step count
• Improved tokenization strategies
• Evaluation on downstream tasks

---

⚠️ Disclaimer

This is research code and an experimental architecture. Results are preliminary and not production-ready.