Update maintained examples for the Ender-60 default

.github/workflows/build-models.yml

@@ -2,34 +2,45 @@ name: Build Example Model Pickles
 
 on:
   workflow_dispatch:
+  pull_request:
+    paths:
+      - ".github/workflows/build-models.yml"
+      - "numerai/**"
   push:
-    # paths:
-    #   - example_model.ipynb
-    #   - hello_numerai.ipynb
-    #   - feature_neutralization.ipynb
-    #   - target_ensemble.ipynb
-    #   - signals/example_model.ipynb
-    #   - crypto/example_model.ipynb
     branches:
       - master
 
-concurrency: build-example-models
+concurrency:
+  group: build-example-models-${{ github.ref }}
+  cancel-in-progress: true
 
 jobs:
 
   build_and_test:
     name: "Build Example Model Pickles"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
-      - name: Install jupyter
+      - name: install-notebook-dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install jupyter
-          pip install -r https://raw.githubusercontent.com/numerai/numerai-predict/refs/heads/master/py3.12/requirements.txt
+          pip install \
+            cloudpickle==3.1.1 \
+            jupyter \
+            lightgbm==4.5.0 \
+            matplotlib==3.10.3 \
+            numerapi==2.20.7 \
+            pandas==2.3.1 \
+            pyarrow==18.1.0 \
+            scikit-learn==1.6.1 \
+            scipy==1.16.0 \
+            seaborn==0.13.2
+          pip install --no-deps numerai-tools==0.4.0
+      - name: test-agent-helpers
+        run: PYTHONPATH=numerai python -m unittest numerai.agents.tests.test_target_transforms -v
       - name: build-example-model
         run: |
           jupyter nbconvert \
@@ -55,37 +66,55 @@ jobs:
             --ExecutePreprocessor.timeout=-1 \
             --to html
       - name: build-signals-example-model
+        if: github.event_name == 'push' || github.event_name == 'workflow_dispatch'
         run: |
           jupyter nbconvert \
             --execute signals/example_model.ipynb \
             --ExecutePreprocessor.timeout=-1 \
             --to html
       - name: build-crypto-example-model
+        if: github.event_name == 'push' || github.event_name == 'workflow_dispatch'
         run: |
           jupyter nbconvert \
             --execute crypto/example_model.ipynb \
             --ExecutePreprocessor.timeout=-1 \
             --to html
+      - name: move-numerai-pickles-to-cached-pickles-dir
+        run: |
+          mkdir -p cached-pickles/
+          mv -f numerai/example_model_v53_ender60.pkl cached-pickles/
+          mv -f numerai/hello_numerai_v53_ender60.pkl cached-pickles/
+          mv -f numerai/feature_neutralization_v53_ender60.pkl cached-pickles/
+          mv -f numerai/target_ensemble_v53_ender60.pkl cached-pickles/
+      - name: generate-publication-candidates-and-provenance
+        run: |
+          python numerai/generate_example_predictions.py \
+            --model cached-pickles/example_model_v53_ender60.pkl \
+            --output-dir generated-example-predictions
+      - name: upload-ender60-validation-evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: v53-ender60-example-model-${{ github.sha }}
+          if-no-files-found: error
+          path: |
+            cached-pickles/*_v53_ender60.pkl
+            generated-example-predictions/*
+            numerai/*.html
+      - name: cache-provenance-for-master
+        if: github.event_name == 'push'
+        run: |
+          cp generated-example-predictions/v53_ender60_provenance.json cached-pickles/
       - name: delete-generated-files
         run: |
-          rm -f numerai/example_model.html
-          rm -f numerai/hello_numerai.html
-          rm -f numerai/feature_neutralization.html
-          rm -f numerai/target_ensemble.html
+          rm -f numerai/*.html
           rm -f signals/example_model.html
           rm -f signals/signals_example_preds.csv
           rm -f signals_example_preds.csv
           rm -f crypto/example_model.html
           rm -f crypto/crypto_example_preds.csv
           rm -f crypto_example_preds.csv
-      - name: move-numerai-pickles-to-cached-pickles-dir
-        run: |
-          mkdir -p cached-pickles/
-          mv -f numerai/example_model.pkl cached-pickles/
-          mv -f numerai/hello_numerai.pkl cached-pickles/
-          mv -f numerai/feature_neutralization.pkl cached-pickles/
-          mv -f numerai/target_ensemble.pkl cached-pickles/
       - name: commit-to-master
+        if: github.event_name == 'push'
         uses: EndBug/add-and-commit@v9
         with:
           add: "cached-pickles/*"

AGENTS.md

@@ -301,5 +301,6 @@ Build datasets with `python -m agents.code.data.build_full_datasets`:
 - **Register repo skills**: `ln -s $PWD/numerai/agents/skills/* ~/.codex/skills/`
 - **Network access required** for MCP operations (Codex CLI may need `--yolo` flag)
 - **Always query Python version** before creating pkl files
-- **BMC (Benchmark Model Contribution)** is the key experiment metric (proxy for MMC), computed vs official `v53_lgbm_ender20` benchmark predictions in `*_benchmark_models.parquet`
+- **BMC (Benchmark Model Contribution)** is the key experiment metric (proxy for MMC), computed by default vs official `v53_lgbm_ender60` benchmark predictions in `*_benchmark_models.parquet`.
+- The v5.3 default research target is explicit `target_ender_60` with a 16-era embargo. Ender-20 work must explicitly select `target_ender_20`; never let an Ender-20-named config read generic `target`.
 - **Only Classic tournament (8)** supports pickle uploads

README.md

@@ -9,9 +9,15 @@ Numerai is quickly developing open-source agent skills for you to use in the tou
 ```
 git clone [email protected]:numerai/example-scripts
 cd example-scripts && curl -sL http://numer.ai/install-mcp.sh | bash
-codex exec --yolo "find the best neural network architecture to predict target ender"
+codex exec --yolo "find the best neural network architecture to predict target_ender_60"
 ```
 
+The maintained v5.3 examples pin `target_ender_60` explicitly. After the
+in-place default-target cutover, generic `target` aliases Ender-60, while
+`target_ender_20` remains available for intentionally reproducing the older
+horizon. Explicit target names and the target-versioned cached model filenames
+prevent an old Ender-20 model from being mistaken for the current example.
+
 ## Notebooks
 We highly recommend getting started with Agents using the above section. But, if you're looking to kill some time on artisan data science, you can check out our tutorial notebooks here as well: 
 

cached-pickles/v53_ender60_provenance.json

@@ -0,0 +1,39 @@
+{
+  "generated_at": "2026-08-06T19:22:31.088355+00:00",
+  "source_commit": "085420354bcd88caa29dfb68964bd14fbbc7522a",
+  "source_notebook": "numerai/example_model.ipynb",
+  "dataset_version": "v5.3",
+  "target_col": "target_ender_60",
+  "model_artifact": "example_model_v53_ender60.pkl",
+  "model_sha256": "a84448a2ff635631d111802398fc65fcba8250f943a729d64bbdf255a7b0cd34",
+  "model_artifacts": {
+    "example_model_v53_ender60.pkl": {
+      "source_notebook": "numerai/example_model.ipynb",
+      "sha256": "a84448a2ff635631d111802398fc65fcba8250f943a729d64bbdf255a7b0cd34"
+    },
+    "feature_neutralization_v53_ender60.pkl": {
+      "source_notebook": "numerai/feature_neutralization.ipynb",
+      "sha256": "8c45f0f1062bb337ebc33b1f1a9497a7e71ccc380ea054bb238b147c85d26230"
+    },
+    "hello_numerai_v53_ender60.pkl": {
+      "source_notebook": "numerai/hello_numerai.ipynb",
+      "sha256": "f94465d7431f6c53130fb7361b4132184539181e74e43e077db220d7ac3ca1a7"
+    },
+    "target_ensemble_v53_ender60.pkl": {
+      "source_notebook": "numerai/target_ensemble.ipynb",
+      "sha256": "fabba10e6701984022e28845f0748252e65ac1ffbdc4d4dcefc141a4061d5775"
+    }
+  },
+  "predictions": {
+    "validation": {
+      "file": "validation_example_preds_v53_ender60.parquet",
+      "rows": 4107040,
+      "sha256": "9587b14138c572fd041c1a7c05f0b36a53a3991f911aca162eae8bae9f82a2b3"
+    },
+    "live": {
+      "file": "live_example_preds_v53_ender60.parquet",
+      "rows": 7034,
+      "sha256": "89b4702157b68bee4bfedd0e21e0b7368274c4f7fc861ccbe0af2853a8d7cb45"
+    }
+  }
+}

numerai/AGENTS.md

@@ -1,6 +1,8 @@
 # Numerai Tournament
 This folder contains examples on how to participate in the Numerai Tournament.
 
+For v5.3, select `target_ender_60` explicitly in maintained training examples. The generic `target` column demonstrates the current default alias only; do not use it when the horizon must remain reproducible.
+
 ## Directory Guide:
 - `agents/`: agentic research framework + training/analysis pipeline (`python -m agents.code.modeling`)
 - `v5.3/` (and other `v*/`): Numerai dataset files (often gitignored locally)

numerai/agents/AGENTS.md

@@ -8,6 +8,8 @@ Always check the agents/skills/ folder for skills that match the user request.
 Run commands from `numerai/` (so `agents` is importable), or from repo root with `PYTHONPATH=numerai`.
 Data is expected to live under `numerai/<data_version>/` (e.g. `numerai/v5.3/`), which is often gitignored locally.
 
+The default v5.3 workflow uses explicit `target_ender_60`, `v53_lgbm_ender60`, and a 16-era embargo. Generic `target` is a mutable dataset alias and must not be used where target identity matters. Keep Ender-20-named configs pinned to `target_ender_20`.
+
 To make these repo skills available to Codex CLI, symlink them into `~/.codex/skills/`:
 `ln -s $PWD/numerai/agents/skills/* ~/.codex/skills/`
 

numerai/agents/baselines/configs/deep_lgbm_ender20_baseline.py

@@ -2,7 +2,7 @@
           'embargo_eras': 13,
           'era_col': 'era',
           'feature_set': 'all',
-          'target_col': 'target'},
+          'target_col': 'target_ender_20'},
  'model': {'x_groups': ['features', 'era', 'benchmark_models'], 'params': {'colsample_bytree': 0.1,
                       'learning_rate': 0.001,
                       'max_depth': 10,

numerai/agents/baselines/configs/deep_lgbm_ender20_baseline_downsampled.py

@@ -4,7 +4,7 @@
         "embargo_eras": 13,
         "era_col": "era",
         "feature_set": "all",
-        "target_col": "target",
+        "target_col": "target_ender_20",
         "id_col": "id",
         "full_data_path": "v5.3/downsampled_full.parquet",
         "benchmark_data_path": "v5.3/downsampled_full_benchmark_models.parquet",

numerai/agents/baselines/configs/deep_lgbm_ender60_baseline.py

@@ -0,0 +1,22 @@
+CONFIG = {'data': {'data_version': 'v5.3',
+          'embargo_eras': 16,
+          'era_col': 'era',
+          'feature_set': 'all',
+          'target_col': 'target_ender_60'},
+ 'model': {'x_groups': ['features', 'era', 'benchmark_models'], 'params': {'colsample_bytree': 0.1,
+                      'learning_rate': 0.001,
+                      'max_depth': 10,
+                      'min_data_in_leaf': 10000,
+                      'n_estimators': 30000,
+                      'n_jobs': 30,
+                      'num_leaves': 1024,
+                      'random_state': 1337},
+           'type': 'LGBMRegressor'},
+ 'output': {'output_dir': 'baselines',
+            'results_name': 'deep_lgbm_ender60_baseline'},
+ 'preprocessing': {'missing_value': 2.0, 'nan_missing_all_twos': False},
+ 'training': {'cv': {'embargo': 16,
+                     'enabled': True,
+                     'min_train_size': 0,
+                     'mode': 'expanding',
+                     'n_splits': 5}}}

numerai/agents/baselines/configs/deep_lgbm_ender60_baseline_downsampled.py

@@ -0,0 +1,40 @@
+CONFIG = {
+    "data": {
+        "data_version": "v5.3",
+        "embargo_eras": 16,
+        "era_col": "era",
+        "feature_set": "all",
+        "target_col": "target_ender_60",
+        "id_col": "id",
+        "full_data_path": "v5.3/downsampled_full.parquet",
+        "benchmark_data_path": "v5.3/downsampled_full_benchmark_models.parquet",
+    },
+    "model": {
+        "x_groups": ["features", "era", "benchmark_models"],
+        "params": {
+            "colsample_bytree": 0.1,
+            "learning_rate": 0.001,
+            "max_depth": 10,
+            "min_data_in_leaf": 10000,
+            "n_estimators": 30000,
+            "n_jobs": 30,
+            "num_leaves": 1024,
+            "random_state": 1337,
+        },
+        "type": "LGBMRegressor",
+    },
+    "output": {
+        "output_dir": "baselines",
+        "results_name": "deep_lgbm_ender60_baseline_downsampled",
+    },
+    "preprocessing": {"missing_value": 2.0, "nan_missing_all_twos": False},
+    "training": {
+        "cv": {
+            "embargo": 16,
+            "enabled": True,
+            "min_train_size": 0,
+            "mode": "expanding",
+            "n_splits": 5,
+        }
+    },
+}

numerai/agents/baselines/configs/small_lgbm_ender20_baseline.py

@@ -2,7 +2,7 @@
           'embargo_eras': 13,
           'era_col': 'era',
           'feature_set': 'medium',
-          'target_col': 'target',
+          'target_col': 'target_ender_20',
           'full_data_path': 'v5.3/downsampled_full.parquet',
           'benchmark_data_path': 'v5.3/downsampled_full_benchmark_models.parquet'},
  'model': {'x_groups': ['features', 'era', 'benchmark_models'], 'params': {'colsample_bytree': 0.1,

numerai/agents/baselines/configs/small_lgbm_ender60_baseline.py

@@ -0,0 +1,25 @@
+CONFIG = {'data': {'data_version': 'v5.3',
+          'embargo_eras': 16,
+          'era_col': 'era',
+          'feature_set': 'medium',
+          'target_col': 'target_ender_60',
+          'full_data_path': 'v5.3/downsampled_full.parquet',
+          'benchmark_data_path': 'v5.3/downsampled_full_benchmark_models.parquet'},
+ 'model': {'x_groups': ['features', 'era', 'benchmark_models'], 'params': {'colsample_bytree': 0.1,
+                      'device_type': 'gpu',
+                      'learning_rate': 0.01,
+                      'max_depth': 5,
+                      'min_data_in_leaf': 10000,
+                      'n_estimators': 2000,
+                      'n_jobs': 30,
+                      'num_leaves': 31,
+                      'random_state': 1337},
+           'type': 'LGBMRegressor'},
+ 'output': {'output_dir': 'baselines',
+            'results_name': 'small_lgbm_ender60_baseline'},
+ 'preprocessing': {'missing_value': 2.0, 'nan_missing_all_twos': False},
+ 'training': {'cv': {'embargo': 16,
+                     'enabled': True,
+                     'min_train_size': 0,
+                     'mode': 'expanding',
+                     'n_splits': 5}}}

numerai/agents/code/analysis/plot_benchmark_corrs.py

@@ -28,23 +28,27 @@
 DEFAULT_DATA_VERSION_V52 = "v5.2"
 DEFAULT_DATA_VERSION_V51 = "v5.1"
 DEFAULT_DATA_VERSION_V5 = "v5.0"
-DEFAULT_MODELS = ["v52_teager20", "v51_teager20", "v5_teager20"]
+DEFAULT_MODELS = ["v53_ender60", "v52_teager20", "v51_teager20", "v5_teager20"]
 
 ALIASES: dict[str, Sequence[str]] = {
     "v52_teager20": ("v52_lgbm_teager2b20", "v52_lgbm_teager20"),
     "v51_teager20": ("v51_lgbm_teager2b20", "v51_lgbm_teager20", "v51_teager20"),
     "v5_teager20": ("v5_lgbm_teager2b20", "v5_lgbm_teager20", "v5_teager20"),
     "v53_ender20": ("v53_lgbm_ender20",),
+    "v53_ender60": ("v53_lgbm_ender60",),
     "v52_ender20": ("v52_lgbm_ender20",),
     "ender20": ("v53_lgbm_ender20",),
+    "ender60": ("v53_lgbm_ender60",),
     "v52_cyrus": ("v52_lgbm_cyrusd20",),
     "cyrus": ("v52_lgbm_cyrusd20",),
     "cyrusd20": ("v52_lgbm_cyrusd20",),
 }
 
 TARGET_ALIASES: dict[str, Sequence[str]] = {
     "ender20": ("target_ender_20", "target_ender20"),
+    "ender60": ("target_ender_60", "target_ender60"),
     "v53_ender20": ("target_ender_20",),
+    "v53_ender60": ("target_ender_60",),
     "v52_ender20": ("target_ender_20",),
     "cyrus": ("target_cyrusd_20", "target_cyrus_20", "target_cyrus20"),
     "cyrusd20": ("target_cyrusd_20",),
@@ -55,7 +59,7 @@
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
         description=(
-            "Plot cumulative per-era numerai_corr for benchmark models vs ender/cyrus."
+            "Plot cumulative per-era numerai_corr for benchmark models vs Ender-60/Cyrus."
         )
     )
     parser.add_argument(
@@ -139,8 +143,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--ender-col",
         type=str,
-        default="ender20",
-        help="Target column (or alias) for ender20 correlations.",
+        default="ender60",
+        help="Target column (or alias) for Ender-60 correlations.",
     )
     parser.add_argument(
         "--cyrus-col",
@@ -223,7 +227,7 @@ def _resolve_target_column(name: str, columns: Iterable[str]) -> str:
 
 def _infer_version(name: str, default_version: str, args: argparse.Namespace) -> str:
     lowered = name.lower()
-    if lowered in {"ender20"}:
+    if lowered in {"ender20", "ender60"}:
         return args.v53_version
     if lowered.startswith("v53_"):
         return args.v53_version