xsong69/stablelm-runs-all

StableLM FineWeb Training Runs

This repository contains multiple StableLM training runs and checkpoints trained on FineWeb-Edu, with different weight decay configurations.

⚠️ This repository is intended for research and analysis purposes. It is not a single ready-to-use Hugging Face model.

---

📁 Repository Structure

stablelm_fineweb_wd0/ stablelm_fineweb_wd0.01/ stablelm_fineweb_wd0.1/ stablelm_fineweb_wd1/ stablelm_fineweb_wd3/

Inside each folder you will find:

• stepXXXX-unsharded/: unsharded model checkpoints
• logs/: training logs
• wandb/: experiment tracking data
• train_data/: metadata for training

---

🧪 Training Setup

• Base architecture: StableLM (1B)
• Dataset: FineWeb-Edu (tokenized)
• Optimizer: AdamW
• Global batch size: 128
• Sequence length: 4096
• Weight decay: varies per run

Each run was trained independently using identical configurations except for the weight decay parameter.

---

📌 Intended Use

This repository is intended for:

• Studying the effect of weight decay on large language model training
• Analyzing optimization dynamics
• Reproducing or extending StableLM training experiments

It is not recommended to directly deploy these checkpoints in production without additional validation and fine-tuning.

---

📥 Downloading Checkpoints

You can download specific runs using:

git lfs install
git clone https://huggingface.co/xsong69/stablelm-runs-all



Each subfolder corresponds to an independent training run: