feat: add autoresearch files

Author: ndrpp

autoresearch/Dockerfile

@@ -0,0 +1,38 @@
+FROM nvidia/cuda:12.8.0-devel-ubuntu22.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV PYTHONUNBUFFERED=1
+
+# System dependencies
+RUN apt-get update && apt-get install -y \
+    python3 python3-dev python3-venv python3-pip \
+    git curl wget \
+    && rm -rf /var/lib/apt/lists/* \
+    && ln -sf /usr/bin/python3 /usr/bin/python
+
+# Install uv for fast package management
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+ENV PATH="/root/.local/bin:$PATH"
+
+# Install Python dependencies (cached unless these lines change)
+RUN uv pip install --system --no-cache-dir \
+    torch==2.9.1 --index-url https://download.pytorch.org/whl/cu128
+
+RUN uv pip install --system --no-cache-dir \
+    vllm \
+    kernels>=0.11.7 \
+    rustbpe>=0.1.0 \
+    tiktoken>=0.11.0 \
+    pyarrow>=21.0.0 \
+    huggingface-hub \
+    requests>=2.32.0 \
+    numpy>=2.2.6
+
+# Copy application code (changes here won't invalidate install layers)
+WORKDIR /app
+COPY . /app
+
+# Ocean platform runs: python /app/data/transformations/algorithm
+RUN mkdir -p /app/data/transformations && ln -sf /app/algo.py /app/data/transformations/algorithm
+
+CMD ["python", "algo.py"]

autoresearch/README.md

@@ -0,0 +1,55 @@
+# Autoresearch on Ocean Network
+
+Autonomous ML research agent that iteratively improves a GPT pretraining script to minimize validation bits-per-byte (val_bpb). Inspired by [Karpathy's autoresearch](https://github.com/karpathy/autoresearch).
+
+The key difference: everything runs **inside a single Docker container** on an [Ocean](https://dashboard.oncompute.ai/) GPU node (H200, 141GB VRAM) with a **local open-source LLM** — no API keys needed.
+
+## How It Works
+
+1. **Data prep** — Downloads HuggingFace data shards, trains a BPE tokenizer (`prepare.py`)
+2. **Load agent LLM** — Qwen3-32B-AWQ via vLLM (~18GB VRAM, stays resident)
+3. **Baseline run** — Runs the original `train.py` (5-min training budget), records val_bpb
+4. **Agent loop** (up to 200 iterations):
+   - LLM reads experiment history + current best `train.py`
+   - Generates a hypothesis + complete new `train.py`
+   - Syntax check → train (5 min) → evaluate val_bpb
+   - If improved: keep. If not: revert to best.
+   - `results.json` saved after every iteration
+
+The user extracts `results["best"]["train_py"]` to get the winning code.
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `algo.py` | Core agent loop — orchestrates LLM inference and training |
+| `train.py` | GPT pretraining script (the file the agent modifies) |
+| `prepare.py` | Data download, tokenizer, dataloader, evaluation (read-only) |
+| `program.md` | Instructions for the agent LLM |
+| `Dockerfile` | Container build (CUDA 12.8, Python, PyTorch, vLLM) |
+| `plot_progress.py` | Generate progress charts from results |
+
+## Usage
+
+1. Go to [dashboard.oncompute.ai](https://dashboard.oncompute.ai/)
+2. Select an **H200 GPU** environment
+3. Configure the job and add payment
+4. Open the **Ocean Orchestrator** in VS Code / your editor
+5. Open this directory in the orchestrator and run the job — the container builds and executes `algo.py` autonomously
+6. Download `results.json` from the outputs when complete
+
+To plot results after a run:
+```bash
+python plot_progress.py path/to/results.json progress.png
+```
+
+## Results
+
+### Qwen3-32B-AWQ — First Run
+
+![Qwen3-32B first run](assets/images/qwen32B_first_run_progress.png)
+
+- **Baseline**: 1.0077 val_bpb
+- **Best**: 0.9818 val_bpb (2.6% improvement)
+- **201 iterations** over 5.5 hours, 30 successful runs (85% crash rate)
+- Key improvements: increased model depth (8→10 layers), late-stage hyperparameter tuning