Add HF Space config frontmatter to README

README.md

@@ -1,3 +1,14 @@
+---
+title: Modular Addition Feature Learning
+emoji: 🔢
+colorFrom: blue
+colorTo: yellow
+sdk: gradio
+sdk_version: "6.5.1"
+app_file: hf_app/app.py
+pinned: false
+---
+
 # On the Mechanism and Dynamics of Modular Addition
 
 ### Fourier Features, Lottery Ticket, and Grokking
@@ -36,14 +47,55 @@ python hf_app/app.py
 
 ### Deploy to Hugging Face Spaces
 
-1. Create a new Space at [huggingface.co/new-space](https://huggingface.co/new-space) (SDK: Gradio)
-2. Push the repo:
-   ```bash
-   git remote add hf https://huggingface.co/spaces/YOUR_USERNAME/YOUR_SPACE_NAME
-   git push hf main
-   ```
-3. The app reads from `precomputed_results/` — the included examples (p=15, 23, 29, 31) work out of the box
-4. Users can generate results for additional $p$ values on-demand via the "Generate" button. New results are auto-committed back to the Space repo so they persist.
+We use the [Hugging Face Python API](https://huggingface.co/docs/huggingface_hub/) to upload to Spaces, since HF now requires [Xet storage](https://huggingface.co/docs/hub/xet) for binary files (PNGs, etc.) which standard `git push` does not handle.
+
+**First-time setup:**
+
+```bash
+pip install huggingface_hub hf_xet
+```
+
+Log in (get a **write** token from https://huggingface.co/settings/tokens):
+
+```bash
+huggingface-cli login
+```
+
+**Upload to the Space:**
+
+```python
+from huggingface_hub import HfApi
+api = HfApi()
+api.upload_folder(
+    folder_path=".",
+    repo_id="y-agent/modular-addition-feature-learning",
+    repo_type="space",
+    ignore_patterns=[
+        "trained_models/*", "saved_models/*", "src/saved_models/*",
+        ".git/*", ".claude/*", ".DS_Store", "tmp/*",
+        "notebooks/*", "figures/*", "__pycache__/*", "src/wandb/*",
+    ],
+    commit_message="Update app",
+)
+```
+
+Or as a one-liner from the project root:
+
+```bash
+python -c "
+from huggingface_hub import HfApi; HfApi().upload_folder(
+    folder_path='.', repo_id='y-agent/modular-addition-feature-learning',
+    repo_type='space', ignore_patterns=[
+        'trained_models/*','saved_models/*','src/saved_models/*',
+        '.git/*','.claude/*','.DS_Store','tmp/*',
+        'notebooks/*','figures/*','__pycache__/*','src/wandb/*'],
+    commit_message='Update app')
+"
+```
+
+**What gets uploaded:** Only the files the app needs — `hf_app/`, `precompute/`, `precomputed_results/`, `src/`, `requirements.txt`, `README.md`. Model checkpoints, notebooks, and figures are excluded.
+
+**On-demand training:** Users can generate results for new $p$ values directly from the app's "Generate" button. Streaming logs show real-time training progress. New results are auto-committed back to the Space repo so they persist across restarts.
 
 > **Tip:** For GPU-accelerated on-demand training, select a GPU runtime in your Space settings.
 

run_experiment.sh

@@ -0,0 +1,52 @@
+#!/bin/bash
+
+#SBATCH --job-name=tk_module_addition_feature # Job name
+#SBATCH --partition=gpu
+#SBATCH --gres=gpu:h100:1
+#SBATCH --qos=qos_zhuoran_yang
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=16
+#SBATCH --time=48:00:00
+#SBATCH --output=slurm_output/%j.out
+#SBATCH --error=slurm_output/%j.err
+#SBATCH --requeue
+
+# Set working directory explicitly
+WORK_DIR=/home/jh3439/modular-addition-feature-learning
+
+echo '-------------------------------'
+cd ${WORK_DIR}
+echo "Working directory: $(pwd)"
+echo Running on host $(hostname)
+echo Time is $(date)
+echo '-------------------------------'
+echo -e '\n\n'
+
+export PROCS=${SLURM_CPUS_ON_NODE}
+
+module load CUDA
+module load cuDNN
+module load miniconda
+
+# Initialize conda for bash - try multiple methods
+source $(conda info --base)/etc/profile.d/conda.sh
+conda activate llm_base
+
+echo "Python path: $(which python)"
+echo "Python version: $(python --version)"
+echo "Conda environment: $CONDA_DEFAULT_ENV"
+
+echo "Starting experiments..."
+echo "============================================================="
+
+cd src
+
+# Use explicit Python path from llm_base environment
+/gpfs/radev/home/jh3439/.conda/envs/llm_base/bin/python module_nn.py --init_type random --act_type ReLU --optimizer AdamW --init_scale 0.1
+#python module_nn.py --init_type random --act_type ReLU --optimizer SGD --lr 0.1 --init_scale 0.01
+#python module_nn.py --init_type single-freq --act_type Quad --optimizer SGD --lr 0.1 --init_scale 0.02
+#python module_nn.py --init_type single-freq --act_type ReLU --optimizer SGD --lr 0.01 --init_scale 0.002
+#python module_nn.py --init_type random --act_type Quad --optimizer SGD --lr 0.1 --init_scale 0.1
+
+
+#python module_nn.py --init_type random --act_type ReLU --optimizer AdamW --init_scale 0.1 --frac_train 0.75 --weight_decay 2 --lr 1e-4 --num_epochs 50000 --d_mlp 128