Update TabArena HF job script

benchmark/tabarena/hf_job.sh

@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+# FelaTab x TabArena on Hugging Face Jobs (CPU).
+# Launched via huggingface_hub HfApi.run_job (see benchmark/tabarena/launch_hf_job.py).
+# Expects: HF_TOKEN in env (job secret), writes results to dataset repo
+# lowdown-labs/fela-tab-tabarena-results.
+set -euxo pipefail
+
+WORK=/workspace
+TABARENA_SUBSET="${TABARENA_SUBSET:-lite}"
+RUN_NAME="${RUN_NAME:-felatab_hfjob}"
+
+mkdir -p "$WORK/tmp" && cd "$WORK"
+
+pip install -q uv
+
+# 1) FelaTab repo (model + this benchmark code) from HF
+python - <<'EOF'
+from huggingface_hub import snapshot_download
+snapshot_download(
+    "lowdown-labs/fela-tab",
+    allow_patterns=[
+        "benchmark/**", "modeling.py", "configuration_felatab.py",
+        "config*.json", "model_small_int8.safetensors", "model_big_int8.safetensors",
+    ],
+    local_dir="/workspace/fela-tab",
+)
+print("fela-tab snapshot ready")
+EOF
+
+# 2) TabArena in its own uv venv (Python 3.12 + pre-release AutoGluon)
+git clone --depth 1 https://github.com/autogluon/tabarena.git
+cd tabarena
+uv venv --seed --python 3.12 .venv
+uv pip install --python .venv/bin/python --prerelease=allow -e "./packages/tabarena[benchmark]"
+uv pip install --python .venv/bin/python torch safetensors tabulate "botocore[crt]"
+
+# 3) Run FelaTab on TabArena
+cd "$WORK/fela-tab/benchmark/tabarena"
+TMPDIR="$WORK/tmp" AWS_CONFIG_FILE=/dev/null AWS_EC2_METADATA_DISABLED=true \
+  "$WORK/tabarena/.venv/bin/python" -u run_tabarena.py \
+  --subset "$TABARENA_SUBSET" --run-name "$RUN_NAME" 2>&1 | tee "$WORK/run.log"
+
+# 4) Upload results
+python - <<EOF
+from huggingface_hub import HfApi
+api = HfApi()
+api.create_repo("lowdown-labs/fela-tab-tabarena-results", repo_type="dataset", exist_ok=True)
+try:
+    api.upload_file(path_or_fileobj="$WORK/run.log", path_in_repo="logs/$RUN_NAME.log",
+                    repo_id="lowdown-labs/fela-tab-tabarena-results", repo_type="dataset")
+    print("uploaded log")
+except Exception as e:
+    print("log upload failed", e)
+try:
+    api.upload_folder(folder_path="$WORK/fela-tab/benchmark/tabarena/eval/$RUN_NAME",
+                      path_in_repo="eval/$RUN_NAME",
+                      repo_id="lowdown-labs/fela-tab-tabarena-results", repo_type="dataset")
+    print("uploaded eval")
+except Exception as e:
+    print("eval upload failed", e)
+EOF
+echo "JOB_DONE"