README: present Docker (harbor) and venv as two explicit run options

README.md

@@ -19,7 +19,13 @@ description + data/eval context + GPU hint + submission format, and **no**
 specialized-library mandate. AutoDS appends its C1 specialized-library layer at
 runtime (`c1_prompt.py`); any other agent uses the base instruction as-is.
 
-## Run
+Task names (`-i`): spaceship-titanic, house-price, amp-parkinsons, imdb,
+feedback, cifar10, ogbn-arxiv, fathomnet, identify-contrails.
+
+## Run — pick one: Docker (harbor) or venv (no Docker)
+
+### Option A — harbor (Docker)
+Standard flow; builds a container per task. Works for all 9 tasks.
 ```bash
 REPO=https://huggingface.co/datasets/danil-e/harbor-datasets-mlab
 # AutoDS (C1 — specialized libraries added at runtime):
@@ -27,10 +33,20 @@ harbor run --repo $REPO -d mlab-real -i feedback \
   -a autods_harbor.agent:AutoDSAgent -m <model>        # or any harbor agent
 # Baseline (AutoDS without C1):  add  --ae AUTODS_C1_DISABLED=1
 ```
-`-i` is a task (spaceship-titanic, house-price, amp-parkinsons, imdb, feedback,
-cifar10, ogbn-arxiv, fathomnet, identify-contrails).
 
-**Full-resolution / GPU runs without docker** (docker-in-docker unavailable):
-use the venv runner in the AutoDS fork —
-`github.com/florentiner/AutoDS-Tools` (branch `mlab-gpu-pub`) → `mlab-gpu/run_venv.sh`.
-It pulls each task from this HF repo via `--repo`.
+### Option B — venv (no Docker)
+For a GPU box that is itself a container (docker-in-docker unavailable) and for
+full-resolution runs. Covers the 4 full-data tasks (amp-parkinsons, feedback,
+fathomnet, identify-contrails) via the AutoDS fork; it pulls each task from this
+HF repo through `--repo` (no manual clone).
+```bash
+git clone https://github.com/florentiner/AutoDS-Tools -b mlab-gpu-pub
+cd AutoDS-Tools/mlab-gpu
+export AUTODS_MODEL=<model> AUTODS_API_KEY=... AUTODS_BASE_URL=https://openrouter.ai/api/v1
+./run_venv.sh setup                 # build venvs (GPU torch by default; TORCH_CPU=1 for CPU)
+./run_venv.sh feedback c1           # AutoDS C1  (base = AutoDS without C1)
+./run_venv.sh all                   # all 4 full-data tasks, baseline + C1
+```
+
+Both options run the same base `instruction.md`; `c1` lets AutoDS append its
+specialized-library layer, baseline sets `AUTODS_C1_DISABLED=1`.