Add kaggle/KAGGLE_SETUP.md

kaggle/KAGGLE_SETUP.md

@@ -0,0 +1,43 @@
+# Kaggle GRPO dispatch
+
+## 1. Notebook settings
+- **Accelerator:** GPU T4 x2
+- **Internet:** On
+- **Language:** Python
+
+## 2. Secrets (Add-ons → Secrets → Attach to notebook)
+| Secret name | Value |
+|-------------|--------|
+| `HF_TOKEN` | same as local `.env` |
+| `WANDB_API_KEY` | same as local `.env` |
+| `LANGFUSE_PUBLIC_KEY` | same as local `.env` |
+| `LANGFUSE_SECRET_KEY` | same as local `.env` |
+| `LANGFUSE_HOST` | `https://cloud.langfuse.com` (optional) |
+
+Kaggle does **not** allow uploading secrets via API from this machine — paste once in the UI.
+
+## 3. Checkpoint / observability contract
+Every `save_steps` (default 5):
+1. Held-out MBPP eval (`n_eval=16`)
+2. Langfuse trace `kaggle-ckpt-eval` + scores (`mean_pass_frac`, …)
+3. W&B `eval/*` metrics
+4. **Then** push LoRA to `vanishingradient/deepagents-rl-coding` revision `grpo-step-{n}`
+
+Baseline eval at step 0 runs before training starts.
+
+## 4. Upload notebook
+1. Open [Kaggle Notebooks](https://www.kaggle.com/code) → New Notebook
+2. Apply settings + secrets above
+3. Upload / paste cells from `notebooks/kaggle_grpo.ipynb`
+4. Run All
+
+Or use CLI after `pip install kaggle` (with `~/.kaggle/kaggle.json`):
+
+```bash
+kaggle kernels push -p notebooks/
+```
+
+## 5. Watch
+- Langfuse: filter `kaggle-ckpt-eval`
+- W&B: project `deepagents-rl-coding`, run `kaggle-grpo-mbpp`
+- HF: https://huggingface.co/vanishingradient/deepagents-rl-coding/tree/main