Fix stale README/registry: single dataset, agent-neutral base instructions, data shipped on HF

README.md

@@ -1,34 +1,29 @@
 # harbor-datasets-tabred
 
-Harbor-format **TabReD** dataset (real data), per the Harbor git-repo guide:
-`registry.json` at root + task dirs. Real TabReD (yandex-research/tabred,
-ICLR 2025) — 8 industry tabular tasks, temporal split, real labels.
+Harbor-format **TabReD** benchmark (yandex-research/tabred, ICLR 2025): 8 real
+industry tabular tasks with a **temporal** train/test split. `registry.json` at
+the root, **one folder per task**.
 
-## Datasets (see `registry.json`)
-- **tabred-real** — 8 tasks, AutoDS C1 prompts + LightAutoML
-- **tabred-real-baseline** — same tasks, bare-agent baseline prompts (ablation pair)
+## Dataset (see `registry.json`)
+- **tabred-real** — 8 tasks. Each `datasets/tabred-real/<task>/` ships the real
+  data (`environment/data/`, the 40k-train / 10k-test temporal slice), a base
+  `instruction.md`, a scorer (`tests/score.py`) and the harbor task config.
 
-## Run
-```bash
-# -d takes the DATASET name from registry.json (NOT the repo name)
-harbor run --repo florentiner/harbor-datasets-tabred \
-  -d tabred-real \
-  -a autods_harbor.agent:AutoDSAgent \
-  -m <model>            # e.g. gemma-4-31b-it
-
-# single task:  add  -i weather
-# baseline set: use  -d tabred-real-baseline
-```
-Any Harbor agent works (pass it via `-a`); no agent-side adapter is needed.
-
-## Data is NOT shipped (Kaggle licensing + size)
-The raw TabReD arrays are competition-licensed and large, so they are **not**
-committed. Before running, regenerate each task's `environment/data/` locally
-with the adapter (needs your own `~/.kaggle/kaggle.json` + accepted rules):
+The `instruction.md` is **agent-neutral** — task description (from the tabred
+datasheet) + data + GPU hint + submission format, and **no** specialized-library
+mandate. The AutoDS agent appends its own C1 layer (LightAutoML + training
+discipline) at runtime (`c1_prompt.py`); any other agent uses the base
+instruction as-is. Data **is** included in this repo.
 
+## Run
 ```bash
-# in the AutoDS-Tools repo
-uv run python -m tabred_real.main \
-  --output-dir <path-to>/harbor-datasets-tabred/datasets/tabred-real \
-  --data-root  ~/tabred-real/data
+REPO=https://huggingface.co/datasets/danil-e/harbor-datasets-tabred
+# AutoDS (C1 — specialized libraries added at runtime):
+harbor run --repo $REPO -d tabred-real -i weather \
+  -a autods_harbor.agent:AutoDSAgent -m <model>        # or any harbor agent
+# Baseline (AutoDS without C1):  add  --ae AUTODS_C1_DISABLED=1
 ```
+`-d` is the DATASET name from `registry.json`; `-i` is a task (weather,
+cooking-time, delivery-eta, maps-routing, sberbank-housing, homesite-insurance,
+ecom-offers, homecredit-default). Regenerate/extend data with the `tabred_real`
+adapter if needed.

datasets/tabred-real/README.md

@@ -1 +1,6 @@
 # autods/tabred-real
+
+One folder per task under this directory. Each task ships real data
+(`<task>/environment/data/`), a base `instruction.md` (agent-neutral) and a
+scorer. See the repo root `README.md` for how to run. AutoDS adds its C1
+specialized-library layer at runtime; any other agent uses the base instruction.