菜单
Guide the user through adding a new dataset to the ModernTSF project. Use when the user wants to register a new dataset, integrate a CSV file or pre-split folder, or wire up a custom data source for benchmarking.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Add a new model to ModernTSF by scaffolding it with `tool/tsf.py new-model`, filling the architecture, and verifying with `tsf smoke`. Use when the user wants to integrate a new PyTorch model, register a custom architecture, or wire a new forecaster into the benchmark pipeline.
Verify model(s) end-to-end by running their smoke configs concurrently and reporting PASS/FAIL via `tool/tsf.py smoke`. Use when the user wants to check that a model (new or existing) trains and produces the right output shape, validate a port, or run a quick CI-style gate over many models.
Understand a model in ModernTSF by reading its README card as progressive disclosure — paper venue/date/arXiv/abstract first, then source only if needed. Use when the user asks "what is model X", "tell me about X", "了解 X 模型", "how does X work", which paper is X from, or wants to compare/learn what a forecaster does before running it.
Add or run a *probabilistic* forecasting model in ModernTSF — one that outputs quantiles or a distribution (prediction intervals / uncertainty) instead of a single point, and is scored with CRPS / WQL / coverage. Use this whenever the task involves prediction intervals, quantiles, uncertainty, a probabilistic or distributional forecast, pinball / quantile loss, Gaussian NLL, or the metrics `crps` / `wql` / `coverage_80` / `width_80` — even if the user only says "predict a range" or "give me uncertainty bands". Do NOT use the plain `add-model` flow for these: it builds a point model. Probabilistic output is a separate `output_type` axis, orthogonal to `task.mode`.
Package a finished run into a TSEval Submission Report (result + agent trajectory + report) and contribute it via a GitHub PR to the TSEval leaderboard repo, via `tsf trace` / `tsf submit`. Use when the user wants to submit results to the TSEval leaderboard, capture an experiment trajectory, or build a leaderboard from submissions.
Diagnose and report a ModernTSF framework defect upstream, with user approval, as a GitHub issue or a small verified PR against Diaugeia/ModernTSF. Use after reproducing a crash in src/ or tool/, wrong output or shapes, a broken config or registry, a doc/CLI mismatch, or another repository defect rather than a problem in the user's data or code.
| name | add-dataset |
| description | Guide the user through adding a new dataset to the ModernTSF project. Use when the user wants to register a new dataset, integrate a CSV file or pre-split folder, or wire up a custom data source for benchmarking. |
First ask the user:
date column + numeric channels), split automatically by the built-in custom loader — config only, no code.train.csv/val.csv/test.csv already in one folder — config only (name = "presplit")._read_data — new source files.tool/convert_traffic.py, then a cauair_st config.uv run python tool/tsf.py new-dataset --name my_csv --pattern custom \
--root-path ./dataset/my_csv --data-path my_csv.csv --target OT
# --pattern presplit → config only (folder must hold train/val/test.csv)
# --pattern single → also generates the loader + schema + DATASET_NAME_MAP entry
Put the data under --root-path and reference the new config from a run config via extends. For single, fill the real loader into src/data/datasets/<name>.py::_read_data (use self._get_borders for the split).
For a traffic bundle, build the node bundle first, then point a cauair_st config at it:
uv run python tool/convert_traffic.py \
--values dataset/metr_la/metr_la.npz --values-key data \
--adj dataset/metr_la/adj_mx.pkl --output-dir dataset/metr_la \
--seq-len 12 --pred-len 12 --add-time --freq-min 5 --splits 0.7,0.1,0.2
uv run modern-tsf --config <run config extending the new dataset>
uv run python tool/tsf.py aggregate-plot --dataset <name>
date column; Pattern B treats it as optional (zero timestamps if absent). The scaler is always fitted on train.features = "S") selects the channel via target.[dataset.params] adj_norm to normalize the adjacency for graph models.Complete TOML templates and per-file code for every pattern: docs/en/add-dataset.md; traffic bundles: docs/en/datasets-traffic.md.