competition-baseline: a GPL-3.0 notebook collection of data-competition baselines

What competition-baseline is for

The repository answers a narrow question: how do you start a data competition when you have never done one? Its README states the intent plainly, that baselines are collected and organised because they are simpler to study than winners' code and more practical for beginners. That is the whole scope. It is not a framework, not a benchmark suite, and not a leaderboard tracker. Each competition gets a folder under competition/, and inside sit notebooks and notes for one or more tracks: data mining on structured tables, computer vision, natural language processing, recommendation, and anomaly detection.

The audience follows from that. A student entering a Tianchi or iFlytek contest, a working engineer who has done production ML but never a scored competition, or an instructor assembling teaching material. The README also points to a competition calendar at coggle.club and to a Gitee mirror for readers who cannot reach GitHub quickly, which tells you the maintainers expect a mainland Chinese audience. If you already know how to build a cross-validated pipeline and read a leaderboard, the value here drops sharply.

How the repository is organised and what a folder contains

The layout is a flat directory tree, one folder per event, named after the competition rather than the technique. Examples visible in the README include competition/DataFountain-CCFBDI-2021, competition/DIGIX2021, competition/2024数字中国创新大赛DCIC/海上风电出力预测赛道, and a folder for the 全球AI攻防挑战赛. Inside those folders the material is Jupyter notebooks, which matches the repository's primary language. There is no shared Python package, no requirements.txt at the root, and no CLI. Data flow is manual: you obtain the competition data yourself from the contest platform, open a notebook, and read or run it against your local copy.

That design has a consequence worth stating. Because nothing is packaged, there is no version pinning across notebooks. Two folders from different years may assume different library versions, and a notebook written for an older PyTorch or PaddlePaddle release may need edits before it runs. The README does not claim otherwise; it presents the folders as shares rather than as software. The one quantitative claim in the README is a pointer to a result, the 2024 DCIC offshore wind power track described as a no-external-data solution scoring 0.0715 on the B leaderboard. That is a link to a folder, not a reproduced benchmark, and I have not run it.

Getting the material onto your machine

There is nothing to install from the repository itself, so the commands are git and Jupyter. Clone it, then open a notebook from the folder for the competition you care about:

git clone https://github.com/datawhalechina/competition-baseline.git cd competition-baseline jupyter notebook competition/DataFountain-CCFBDI-2021

If GitHub is slow from your location, the README gives a mirror: git clone https://gitee.com/coggle/competition-baseline.git. After that, the real work is outside the repository. Each notebook expects the competition's own dataset, which you download from the contest platform linked in the README or in the folder notes. There are no config keys, no environment variables and no entry points to document, because the project defines none. Treat each notebook as a script you adapt: check the import block first, because that is where version assumptions live, and check the first data-loading cell, because that is where the expected file paths and column names are encoded.

Where the baseline approach breaks down

A baseline is deliberately not a winning solution, and the README says the quiet part out loud when it contrasts baseline code with winners' code. The practical failure modes follow. First, coverage is uneven. Some events get a single notebook, others get a folder of tracks; the README lists eleven CCF BDCI 2021 tracks in one line each, which suggests breadth over depth in that section. Second, competition data often cannot be redistributed, so the notebooks reference datasets you must fetch from platforms that may have closed registration or removed the files. A notebook whose data link is dead is a reading exercise, not a runnable one. Third, baselines date. A 2021 text-matching notebook predates the current default of fine-tuning a pretrained language model, and copying its feature engineering into a 2026 contest will not place well.

The wrong-tool case is concrete: if your goal is a competitive score, this repository gives you a floor, not a path to the ceiling. If your goal is a maintained dependency you can import, it offers nothing at all.

How it differs from Kaggle notebooks and from framework tutorials

The nearest alternatives are Kaggle's public notebook ecosystem and the official tutorials shipped by frameworks such as PaddlePaddle or MindSpore. Kaggle notebooks are tied to a hosted runtime with the competition data already mounted, and they are searchable by vote count, which is useful when you want many attempts at the same problem. The trade-off is that they live on Kaggle and assume Kaggle's environment. The repository here is the opposite: everything is a file in git, readable offline, forkable, and usable when the competition runs on Tianchi, DataFountain, AIWIN or an iFlytek platform rather than Kaggle. That matters because a large share of the events listed in the README are Chinese platforms, and their data and rules do not travel to Kaggle.

Framework tutorials differ in another direction. A PaddlePaddle or MindSpore tutorial teaches the framework through a clean synthetic task. A folder here teaches a competition through its actual quirks: the metric, the submission format, the class imbalance, the leakage risk. If you want to learn an API, read the framework docs. If you want to learn what a specific contest actually demanded, read the notebook.

Maintenance, licensing and the cost of keeping a notebook archive

The repository is not archived, the default branch is master, and the last push recorded is 2026-07-22. No GitHub releases were retrieved, which is consistent with a project that has no artifact to release. The README is actively extended: it carries sections for the 2026 iFlytek AI developer competition, 2025, 2024 and earlier years, plus the 外滩大会 Deepfake challenge and the AIWIN autumn contests. So the maintenance model is additive. New competitions get new folders; old folders are generally left as they were written. That is cheap to maintain and expensive to consume, because the reader carries the burden of judging whether a 2021 notebook still reflects current practice.

The licence is GPL-3.0. For a notebook collection this is mostly a redistribution question: if you copy a notebook into your own project and distribute that project, the copyleft terms attach to the derivative work. Whether that matters depends on how you reuse the code, and it is worth checking with whoever handles licensing at your organisation rather than assuming. Separately, the competition datasets referenced by the notebooks carry their own terms from the hosting platforms, and those terms are independent of the repository licence.

A concrete way to use it without wasting a week

Pick one folder whose competition matches your current problem type, not the newest folder in the tree. Open the notebook and read the evaluation section first, since the metric determines everything downstream. Then check whether the data link in the README or the folder notes still resolves; if it does not, move to another folder rather than reconstructing the dataset. If it does, run the notebook end to end once with the environment it expects, and treat any failure as information about how old the code is. The 2024 DCIC offshore wind folder is a reasonable test case because the README attaches a specific B-leaderboard figure of 0.0715 to a no-external-data approach, which gives you a number to compare against if you reproduce it. That comparison is the only honest way to decide whether a given baseline is still worth your time.

Editorial conclusion

Adopt it if you are preparing for a Chinese data competition and want a readable starting notebook rather than a winning solution, or if you teach applied machine learning and need worked examples. Skip it if you need a maintained package, a leaderboard-grade pipeline, or per-competition support: the repository is a notebook archive with no releases and no installable artifact, and many linked competitions have already closed. Before relying on any folder, open the notebook, confirm the data source is still reachable, and check that the licence terms of the competition data allow your use.

Official sources

1. datawhalechina/competition-baseline on GitHub
2. Issues
3. License: GPL-3.0
4. Project website
5. README