Introduction | The model has climbed to the top of the leaderboard, and what follows is the most exhausting step: the competition requires submitting a complete technical report covering the model architecture, data augmentation strategy, loss function, and model ensemble plan. You are still tuning the model and have no time to write a long report.
1. The hard part of a competition report isn't "writing" but "recounting"
Competitors are familiar with this scenario: as the competition nears its end, the organizers require a technical report. The report must clearly explain four things—model architecture, data augmentation, loss function, and ensemble plan.
Where is the difficulty? The difficulty is that you produced these four things through trial and adjustment:
• The architecture was revised three times—which version did you end up using?
• You tried seven or eight data augmentation methods—which ones were kept?
• The ensemble plan was put together on the fly during the leaderboard sprint—how do you write it as a clear paragraph?
It's not that you can't write; you need to pull the decisions scattered across months back into a coherent thread.
2. Step 1: Gather materials in Co-Research Workshop
MuchSci's Co-Research Workshop supports team collaboration and can import competition datasets. For competitions, this step is very practical.
| Stage | Why collaboration is needed | What Co-Research Workshop handles |
| Dataset import | Competition data is scattered among teammates | Unified import, available in one place |
| Solution records | Architecture, augmentation, and loss are each changed separately | The team shares the same materials |
| Report writing | One person can't finish it | Multiple people collaborate |
Gather the datasets and key experiment records into place so you have something to base the report on later.
3. Step 2: Let Paper Sailor generate a complete technical report
Once the materials are ready, use Paper Sailor to generate the technical report. It organizes sections according to the structure of a competition report: problem background, data processing, model architecture, training strategy (data augmentation and loss function), ensemble plan, and results and analysis.
The key point: every number in the report comes from your experiment records, and it will not fabricate experimental metrics—this is especially important in competition reports because the organizers will verify your results.
4. Step 3: Draw model architecture diagrams and performance comparison charts
The most time-consuming part of a technical report is actually the figures:
• Model architecture diagram: drawing the network as a flowchart is much more efficient than describing it in text.
• Performance comparison chart: score comparisons across different approaches and versions, explained at once with grouped bar charts.
And the output is journal-size vector SVG, with editable text and no distortion when enlarged—clear even when pasted into an appendix.
5. Step 4: Add references
Competition reports also need citations: where the backbone comes from, which paper the data augmentation strategy is from, and which paper is the original source of the ensemble method. Paper Sailor automatically manages these citations, uses cite markers in the text, generates the bib with one click, and runs an authenticity check before submission.
6. A complete walkthrough
Take an image competition as an example:
1. Create a new project in Co-Research Workshop and import the competition dataset.
2. Gather materials: organize the final architecture, augmentation strategy, loss, and ensemble plan into it.
3. Generate the technical report: produce a draft according to the structure required by the competition.
4. Produce figures: model architecture flowchart + score comparison chart of versions (vector, journal-size).
5. Check numbers: make sure every metric in the report matches the experimental results.
6. Add citations and run the authenticity check.
7. Export and submit according to the organizers' requirements.
7. Common pitfalls
• The report is inconsistent with the submitted version. The architecture described in the report is not the version you finally submitted, and the judges will spot it as soon as they compare.
• Only writing "we used X" without "why we used it". The value of a technical report lies in the rationale for decisions.
• Relying on screenshots for architecture diagrams. Bitmaps blur when enlarged; use vector flowcharts.
• Only reporting the highest score. Without multi-version comparisons, your tuning path isn't visible.
8. Conclusion
A competition technical report is the work of "recounting months of tuning", and the most exhausting part has never been writing ability. MuchSci's Co-Research Workshop gathers the materials, and Paper Sailor handles drafting, figures, and citations—what you need to do is explain the final version of the solution clearly.
Pre-submission checklist:
• Are the architecture and strategy in the report consistent with the final submitted version?
• Does every number match the experiment records?
• Is the architecture diagram vector-based, and is there a performance comparison chart?
• Are all citations real and verifiable?