🧪 Latest features land on the dev branch first. main is the stable, versioned release line — if you want the newest (pre-release) capabilities, track or install from dev.

---

🧭 What is ModernTSF

You don't build a car to drive one, mill flour to bake a loaf, or grow your own beans for a cup of coffee — you reach for something ready-made. AI research needs the same layer: today's agents can write code and run experiments, yet most of the effort — human and agent alike — still goes into reproducing prior work, validating baselines, debugging environments, and writing glue code. ModernTSF is that missing infrastructure layer for time-series forecasting. You bring the idea; the substrate handles everything around it.

---

✨ Highlights

• 🧠 178 model/method entries, 80 dataset presets — a flat catalog spanning baselines, neural forecasters, graph models, custom CSVs, traffic graphs, and GIFT-EVAL
• 🤖 Agent-ready — open the repository in Codex, Claude Code, Pi, or DeepSeek Harness and request a complete workflow in plain language
• 🎛️ Three data settings — time_series, spatiotemporal, and covariate, switchable per run
• 🔁 Reproducible & auditable — TOML configs, fixed seeds, profiled outputs, and optional research rounds keep results comparable without burdening one-off runs
• 🛠️ One entry point — tsf scaffolds, smoke-tests, sweeps, aggregates, ranks, plots, and reports

---

🏁 How to use

git clone https://github.com/Diaugeia/ModernTSF.git
cd ModernTSF
codex

Then just say what you want, in plain language:

> Set up the environment for my GPU.
> Benchmark DLinear, PatchTST and iTransformer on ETTh1 and give me a leaderboard.
> Here is my CSV of hourly sales — add it as a dataset and find the best model for it.
> I have an idea: <describe it>. Scaffold a model, implement it, and compare it against strong baselines.

The Agent owns research planning, code/config editing, diagnosis, interpretation, and reporting through its native capabilities. ModernTSF supplies callable contracts, evidence, budgets, resource leases, and recoverable execution. Python APIs and the optional CLI expose the same services; no extra Agent loop is required.

The wheel includes read-only model cards, configs, verification evidence, components, skills, and task harnesses, so catalog and Agent discovery also work after installation. Use a git checkout for commands that add or rewrite models, datasets, documentation, or verification evidence.

For direct discovery, the public catalogs are equally lightweight:

uv run tsf model list --details
uv run tsf component list
uv run tsf agent task list
uv run tsf agent task render autoresearch --set 'question=<your question>'
uv run tsf agent task start autoresearch --set 'question=<your question>' --json

task start prepares the bounded research round and prompt; it does not dispatch another Agent process.

Optional execution controls add environment audits, TensorBoard/W&B, resource budgets, GPU scheduling, and epoch-boundary recovery without changing the basic run command.

The workflow guide explains the model interface, shared components, offline official Foundation runtimes, data layers, verification, and experiments.

Dataset resources have three deliberately separate layers: ignored local files live under dataset/, executable loaders and schemas live under src/data/, and readable catalog cards live under catalog/datasets/. Code and cards never embed local dataset payloads.

---

📖 Documentation

The compact workflow reference covers models, data, verification, and experiments; exact command options stay in CLI help:

Workflow documentation

But chances are you'll never need any of this — let the agent do the reading.

---

📜 License

ModernTSF is released under the MIT License — open by default, free to use, modify, and build upon.

Copyright © 2026 Diaugeia.AI.

Ordinary paper architectures are maintained locally under the project license. Released pretrained Foundation Models use optional official packages and unchanged checkpoints through the offline runtime boundary; see THIRD_PARTY_NOTICES.md for dependency attribution.

---

⭐ Star History