Self-contained task images (FROM python:3.12-slim, own requirements.txt); remove external base-image and agent-specific coupling

datasets/mlab-real/README.md

@@ -1,6 +1,6 @@
-# autods/mlab-real
+# mlab-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.
+One folder per task. Each task is self-contained: real data
+(`<task>/environment/data/`), a self-contained `environment/Dockerfile`
+(builds from `python:3.12-slim`, no external base image), an agent-neutral
+`instruction.md`, and a scorer (`tests/`). See the repo root README to run.

datasets/mlab-real/amp-parkinsons/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-tabular:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/tabular/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/amp-parkinsons/environment/requirements.txt

@@ -0,0 +1,7 @@
+# Tabular stack (CPU-friendly). torch is installed separately in the Dockerfile.
+lightautoml
+featuretools
+feature-engine
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/amp-parkinsons/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/amp-parkinsons"
+name = "mlab/amp-parkinsons"
 description = "MLAgentBench amp-parkinsons — full competition data (SMAPE)"
-keywords = ["mlab", "mlagentbench", "tabular", "autods", "full-data"]
+keywords = ["mlab", "mlagentbench", "tabular", "full-data"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "tabular"
 metric = "SMAPE"
-specialized_libraries = ["lightautoml", "tsfresh"]
 
 [verifier]
 timeout_sec = 1800.0

datasets/mlab-real/amp-parkinsons/tests/test.sh

@@ -1,7 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/tabular/bin/python
-[ -x "$PY" ] || PY=python
+PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/cifar10/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-vision:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" "torchvision==0.17.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" "torchvision==0.17.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/vision/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/cifar10/environment/requirements.txt

@@ -0,0 +1,9 @@
+# Vision stack. torch + torchvision are installed separately in the Dockerfile.
+timm
+albumentations
+segmentation-models-pytorch
+opencv-python-headless
+pillow
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/cifar10/solution/solve.sh

@@ -2,7 +2,7 @@
 # Oracle: torchvision.models CNN (resnet18 adapted to 32x32) — the specialized
 # library this task names.
 set -euo pipefail
-/opt/venvs/vision/bin/python - <<'PY'
+python - <<'PY'
 import numpy as np, pandas as pd, torch, torch.nn as nn
 from torchvision import models
 

datasets/mlab-real/cifar10/task.toml

@@ -1,15 +1,14 @@
 schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 [task]
-name = "autods-mlab/cifar10"
+name = "mlab/cifar10"
 description = "MLAgentBench CIFAR-10 (small) — image classification (timm / torchvision)"
-keywords = ["mlab", "mlagentbench", "image", "vision", "timm", "torchvision", "autods"]
+keywords = ["mlab", "mlagentbench", "image", "vision", "torchvision"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 [metadata]
 family = "vision"
 metric = "accuracy"
-specialized_libraries = ["timm", "torchvision", "albumentations"]
 [verifier]
 timeout_sec = 1200.0
 collect = []

datasets/mlab-real/cifar10/tests/test.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/vision/bin/python; [ -x "$PY" ] || PY=python
+PY=python; [ -x "$PY" ] || PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/dataset.toml

@@ -1,66 +1,41 @@
-# Dataset manifest for autods/mlab-real
-# Add tasks using: harbor add <org>/<name>
-# Publish using: harbor publish
+# Dataset manifest for mlab-real
 
 [dataset]
-name = "autods/mlab-real"
-description = ""
-authors = []
-keywords = []
+name = "mlab-real"
+version = "2.0"
 
 [[tasks]]
-name = "autods-mlab/amp-parkinsons"
-digest = "sha256:f73327bdbc3611887ebc7320de1b60b6830181f835f43c4aec1272c1b431135b"
+name = "mlab/amp-parkinsons"
+path = "amp-parkinsons"
 
 [[tasks]]
-name = "autods-mlab/amp-parkinsons-base"
-digest = "sha256:ccb6a6c714ff89274aff9c1c9fbebc70bf4fd65dbf8623bf00ad839a4d96cc65"
+name = "mlab/cifar10"
+path = "cifar10"
 
 [[tasks]]
-name = "autods-mlab/cifar10"
-digest = "sha256:82338bbcca781c0f85978d376d3e6a3976f0de3626c4d72ec251251c848c7ef7"
+name = "mlab/fathomnet"
+path = "fathomnet"
 
 [[tasks]]
-name = "autods-mlab/cifar10-base"
-digest = "sha256:794f2afe4575cd7cb0441ffb0dc357a6dc89968d8da4f04d117158694ebb4799"
+name = "mlab/feedback"
+path = "feedback"
 
 [[tasks]]
-name = "autods-mlab/feedback"
-digest = "sha256:9bb6e898f33a7444ea670deeef3e1704ed2972441dbee5e132f3312ec967a637"
+name = "mlab/house-price"
+path = "house-price"
 
 [[tasks]]
-name = "autods-mlab/feedback-base"
-digest = "sha256:8638bdd776e41385cb6be404619d59b14007a1f7b8a7228f1de7f4f63c80923b"
+name = "mlab/identify-contrails"
+path = "identify-contrails"
 
 [[tasks]]
-name = "autods-mlab/house-price"
-digest = "sha256:1d24e80759d8ec3490bc1bfe4906ef03ab1a1ef4f8b87999b06286c4b8dd6b9d"
+name = "mlab/imdb"
+path = "imdb"
 
 [[tasks]]
-name = "autods-mlab/house-price-base"
-digest = "sha256:121c15cbe344c17b65ed6aa5f254e6f28971d6c575bbc701644c22ca21cdadf8"
+name = "mlab/ogbn-arxiv"
+path = "ogbn-arxiv"
 
 [[tasks]]
-name = "autods-mlab/imdb"
-digest = "sha256:5223ed84d3f38ddd8b70aa4a73248c1e9c6489b86f7fde7a274814eb9c42f6f1"
-
-[[tasks]]
-name = "autods-mlab/imdb-base"
-digest = "sha256:239cbaffb8ccba5fcd937f95ab0473f1addfd8756171c67231a2bd12f88e112d"
-
-[[tasks]]
-name = "autods-mlab/ogbn-arxiv"
-digest = "sha256:f7549278a351b19f37d46bf9f50498e377a399ca01495564c6274502d3031b59"
-
-[[tasks]]
-name = "autods-mlab/ogbn-arxiv-base"
-digest = "sha256:0d3ec43a85c2ebad2809ef0590ac76131e936fd6d83caf4dd0793e740e842f8d"
-
-[[tasks]]
-name = "autods-mlab/spaceship-titanic"
-digest = "sha256:3dbc9593b446adf483db1e39a31b5d2fecb27083424a302d0c329c8634737f1d"
-
-[[tasks]]
-name = "autods-mlab/spaceship-titanic-base"
-digest = "sha256:b704bd248e6b143d1e594754d2b5200d5a51b9582325267aee0b0a392c3120f0"
-
+name = "mlab/spaceship-titanic"
+path = "spaceship-titanic"

datasets/mlab-real/fathomnet/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-vision:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" "torchvision==0.17.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" "torchvision==0.17.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/vision/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/fathomnet/environment/requirements.txt

@@ -0,0 +1,9 @@
+# Vision stack. torch + torchvision are installed separately in the Dockerfile.
+timm
+albumentations
+segmentation-models-pytorch
+opencv-python-headless
+pillow
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/fathomnet/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/fathomnet"
+name = "mlab/fathomnet"
 description = "MLAgentBench fathomnet — full competition data (micro-F1)"
-keywords = ["mlab", "mlagentbench", "vision", "autods", "full-data"]
+keywords = ["mlab", "mlagentbench", "vision", "full-data"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "vision"
 metric = "micro-F1"
-specialized_libraries = ["timm", "torchvision"]
 
 [verifier]
 timeout_sec = 1800.0

datasets/mlab-real/fathomnet/tests/test.sh

@@ -1,7 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/vision/bin/python
-[ -x "$PY" ] || PY=python
+PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/feedback/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-nlp:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/nlp/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/feedback/environment/requirements.txt

@@ -0,0 +1,8 @@
+# NLP stack. torch is installed separately in the Dockerfile.
+transformers==4.44.2
+sentence-transformers
+datasets
+textstat
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/feedback/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/feedback"
+name = "mlab/feedback"
 description = "MLAgentBench feedback — full competition data (MCRMSE)"
-keywords = ["mlab", "mlagentbench", "nlp", "autods", "full-data"]
+keywords = ["mlab", "mlagentbench", "nlp", "full-data"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "nlp"
 metric = "MCRMSE"
-specialized_libraries = ["transformers", "sentence-transformers"]
 
 [verifier]
 timeout_sec = 1800.0

datasets/mlab-real/feedback/tests/test.sh

@@ -1,7 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/nlp/bin/python
-[ -x "$PY" ] || PY=python
+PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/house-price/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-tabular:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/tabular/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/house-price/environment/requirements.txt

@@ -0,0 +1,7 @@
+# Tabular stack (CPU-friendly). torch is installed separately in the Dockerfile.
+lightautoml
+featuretools
+feature-engine
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/house-price/solution/solve.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 # Oracle: LightAutoML TabularAutoML — the specialized library this task names.
 set -euo pipefail
-/opt/venvs/tabular/bin/python - <<'PY'
+python - <<'PY'
 import numpy as np, pandas as pd
 from lightautoml.automl.presets.tabular_presets import TabularAutoML
 from lightautoml.tasks import Task

datasets/mlab-real/house-price/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/house-price"
+name = "mlab/house-price"
 description = "MLAgentBench House Prices — tabular regression (LightAutoML)"
-keywords = ["mlab", "mlagentbench", "tabular", "regression", "lightautoml", "autods"]
+keywords = ["mlab", "mlagentbench", "tabular", "regression"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "tabular"
 metric = "r2"
-specialized_libraries = ["lightautoml", "featuretools", "feature-engine"]
 
 [verifier]
 timeout_sec = 1200.0

datasets/mlab-real/house-price/tests/test.sh

@@ -3,8 +3,7 @@
 set -uo pipefail
 
 mkdir -p /logs/verifier
-PY=/opt/venvs/tabular/bin/python
-[ -x "$PY" ] || PY=python
+PY=python
 
 "$PY" /tests/score.py \
   --submission /workspace/submission.csv \

datasets/mlab-real/identify-contrails/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-vision:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" "torchvision==0.17.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" "torchvision==0.17.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/vision/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/identify-contrails/environment/requirements.txt

@@ -0,0 +1,9 @@
+# Vision stack. torch + torchvision are installed separately in the Dockerfile.
+timm
+albumentations
+segmentation-models-pytorch
+opencv-python-headless
+pillow
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/identify-contrails/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/identify-contrails"
+name = "mlab/identify-contrails"
 description = "MLAgentBench identify-contrails — full competition data (Dice)"
-keywords = ["mlab", "mlagentbench", "vision", "autods", "full-data"]
+keywords = ["mlab", "mlagentbench", "vision", "full-data"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "vision"
 metric = "Dice"
-specialized_libraries = ["segmentation-models-pytorch"]
 
 [verifier]
 timeout_sec = 1800.0

datasets/mlab-real/identify-contrails/tests/test.sh

@@ -1,7 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/vision/bin/python
-[ -x "$PY" ] || PY=python
+PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/imdb/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-nlp:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/nlp/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/imdb/environment/requirements.txt

@@ -0,0 +1,8 @@
+# NLP stack. torch is installed separately in the Dockerfile.
+transformers==4.44.2
+sentence-transformers
+datasets
+textstat
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/imdb/solution/solve.sh

@@ -2,7 +2,7 @@
 # Oracle: sentence-transformers embeddings + a light classifier head — the
 # specialized library this task names.
 set -euo pipefail
-/opt/venvs/nlp/bin/python - <<'PY'
+python - <<'PY'
 import pandas as pd
 from sentence_transformers import SentenceTransformer
 from sklearn.linear_model import LogisticRegression

datasets/mlab-real/imdb/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/imdb"
+name = "mlab/imdb"
 description = "MLAgentBench IMDb — text sentiment classification (transformers / sentence-transformers)"
-keywords = ["mlab", "mlagentbench", "text", "nlp", "transformers", "autods"]
+keywords = ["mlab", "mlagentbench", "text", "nlp", "transformers"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "nlp"
 metric = "accuracy"
-specialized_libraries = ["transformers", "sentence-transformers", "setfit"]
 
 [verifier]
 timeout_sec = 1200.0

datasets/mlab-real/imdb/tests/test.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/nlp/bin/python; [ -x "$PY" ] || PY=python
+PY=python; [ -x "$PY" ] || PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/ogbn-arxiv/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-graph:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/graph/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/ogbn-arxiv/environment/requirements.txt

@@ -0,0 +1,6 @@
+# Graph stack. torch is installed separately in the Dockerfile.
+torch-geometric
+ogb
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/ogbn-arxiv/solution/solve.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 # Oracle: torch-geometric GraphSAGE — the specialized library this task names.
 set -euo pipefail
-/opt/venvs/graph/bin/python - <<'PY'
+python - <<'PY'
 import numpy as np, pandas as pd, torch, torch.nn.functional as F
 from torch_geometric.nn import SAGEConv
 

datasets/mlab-real/ogbn-arxiv/task.toml

@@ -1,15 +1,14 @@
 schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 [task]
-name = "autods-mlab/ogbn-arxiv"
+name = "mlab/ogbn-arxiv"
 description = "MLAgentBench OGBN-Arxiv (small) — graph node classification (torch-geometric)"
-keywords = ["mlab", "mlagentbench", "graph", "gnn", "torch-geometric", "ogb", "autods"]
+keywords = ["mlab", "mlagentbench", "graph", "gnn", "torch-geometric", "ogb"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 [metadata]
 family = "graph"
 metric = "accuracy"
-specialized_libraries = ["torch-geometric", "ogb"]
 [verifier]
 timeout_sec = 1200.0
 collect = []

datasets/mlab-real/ogbn-arxiv/tests/test.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -uo pipefail
 mkdir -p /logs/verifier
-PY=/opt/venvs/graph/bin/python; [ -x "$PY" ] || PY=python
+PY=python; [ -x "$PY" ] || PY=python
 "$PY" /tests/score.py --submission /workspace/submission.csv --answer /opt/mlab/answer.csv --reward-json /logs/verifier/reward.json
 exit 0

datasets/mlab-real/spaceship-titanic/environment/Dockerfile

@@ -1,5 +1,24 @@
-FROM autods-mlab-tabular:latest
+# Self-contained task image: no external base image, no private repos.
+# Build context = this `environment/` directory (Harbor does this automatically).
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Tooling that agents commonly need to install themselves / fetch resources.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates build-essential ripgrep unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt /tmp/requirements.txt
+# CPU wheels first (small); fall back to the default index on other platforms.
+RUN pip install --upgrade pip wheel "setuptools<81" \
+    && ( pip install "torch==2.2.2" --index-url https://download.pytorch.org/whl/cpu \
+         || pip install "torch==2.2.2" ) \
+    && pip install -r /tmp/requirements.txt
+
+# Stage the task data into /workspace and the answer key outside it.
 COPY prepare.py /opt/mlab/prepare.py
 COPY data /opt/mlab/data
-RUN /opt/venvs/tabular/bin/python /opt/mlab/prepare.py || python /opt/mlab/prepare.py
+RUN python /opt/mlab/prepare.py
+
 WORKDIR /workspace

datasets/mlab-real/spaceship-titanic/environment/requirements.txt

@@ -0,0 +1,7 @@
+# Tabular stack (CPU-friendly). torch is installed separately in the Dockerfile.
+lightautoml
+featuretools
+feature-engine
+scikit-learn
+pandas
+numpy<2

datasets/mlab-real/spaceship-titanic/solution/solve.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 # Oracle: LightAutoML TabularAutoML — the specialized library this task names.
 set -euo pipefail
-/opt/venvs/tabular/bin/python - <<'PY'
+python - <<'PY'
 import numpy as np, pandas as pd
 from lightautoml.automl.presets.tabular_presets import TabularAutoML
 from lightautoml.tasks import Task

datasets/mlab-real/spaceship-titanic/task.toml

@@ -2,16 +2,15 @@ schema_version = "1.3"
 artifacts = ["/workspace/submission.csv"]
 
 [task]
-name = "autods-mlab/spaceship-titanic"
+name = "mlab/spaceship-titanic"
 description = "MLAgentBench Spaceship Titanic — tabular binary classification (LightAutoML)"
-keywords = ["mlab", "mlagentbench", "tabular", "classification", "lightautoml", "autods"]
+keywords = ["mlab", "mlagentbench", "tabular", "classification"]
 [[task.authors]]
-name = "AutoDS"
+name = "mlab"
 
 [metadata]
 family = "tabular"
 metric = "accuracy"
-specialized_libraries = ["lightautoml", "featuretools", "feature-engine"]
 
 [verifier]
 timeout_sec = 1200.0

datasets/mlab-real/spaceship-titanic/tests/test.sh

@@ -1,12 +1,11 @@
 #!/bin/bash
 # Harbor verifier: score /workspace/submission.csv against the held-out answer
-# key and write the reward. Runs in the task container (FROM autods-mlab-tabular),
+# key and write the reward. Runs in the task container,
 # so the tabular child venv's Python (with pandas) is available.
 set -uo pipefail
 
 mkdir -p /logs/verifier
-PY=/opt/venvs/tabular/bin/python
-[ -x "$PY" ] || PY=python
+PY=python
 
 "$PY" /tests/score.py \
   --submission /workspace/submission.csv \