ADTC 2026 — Submission Template

This is the official template repository for the Africa Deep Tech Challenge 2026 Laptop LLM track.

Fork this repository, fill in the required files, and submit your repository URL via adtc-2026.devpost.com.

---

✅ Submission Checklist

Before submitting, confirm every item:

• Your repository is public on GitHub
• metadata.json is fully filled in — no placeholder values remain
• metadata.json contains exactly 2 test prompts in the test_prompts array, written for your chosen domain
• download_model.sh successfully downloads your model to model/
• The downloaded file is a valid GGUF format (.gguf) weight file
• model/*.gguf is listed in .gitignore — do not commit large weight files
• REPORT.md is filled in with your technical writeup
• Running bash download_model.sh completes without errors
• Your model runs entirely offline — zero external network calls during inference

---

📁 Required File Structure

your-submission/
├── metadata.json          ← Required. Team, model, and test prompt metadata.
├── download_model.sh      ← Required. Downloads your .gguf model weight file.
├── REPORT.md              ← Required. Technical writeup (problem, design, benchmarks).
├── model/
│   └── your-model.gguf   ← Downloaded by the script above. Do NOT commit.
└── .gitignore             ← Must exclude *.gguf and model/ from version control.

---

📝 metadata.json

Fill in every field. No field should remain at its placeholder value.

{
  "team_id": "your-team-id",
  "domain": "coding_assistants",
  "language_scope": ["en"],
  "african_alpha_claim": false,
  "budget_laptop_claim": true,
  "submitter": {
    "name": "your-name",
    "email": "your-email@domain.com",
    "github_handle": "your-github"
  },
  "cross_disciplinary_pairing": {
    "discipline": "education",
    "load_bearing": true,
    "description": "Brief description of how your model serves a real-world domain."
  },
  "test_prompts": [
    {
      "prompt_id": "tp_001",
      "prompt": "Your first test prompt, written for your chosen domain."
    },
    {
      "prompt_id": "tp_002",
      "prompt": "Your second test prompt, written for your chosen domain."
    }
  ],
  "model": {
    "name": "YourModel-Q4_K_M",
    "runtime": "llama.cpp",
    "quantization": "GGUF Q4_K_M",
    "parameters_estimate": "1.1B",
    "packaging": "binary_bundle"
  },
  "_runtime": {
    "model_path": "model/your-model.gguf"
  }
}

Field Reference

Field	Required	Description
team_id	✅	Your unique team ID as registered on the ADTF portal
domain	✅	Your challenge track. One of: math_scientific_reasoning, healthcare_medical, agriculture, creative_writing, coding_assistants, corporate_enterprise, autonomous_ai_agents
language_scope	✅	Array of BCP-47 language codes. Must include at least one.
african_alpha_claim	✅	true only if claiming the African Use Case Bonus
budget_laptop_claim	✅	Must be true — all submissions target the 8 GB RAM laptop profile
submitter.name	✅	Full name of the team member submitting the run
submitter.email	✅	Valid email address linked to the registered team
submitter.github_handle	✅	Verifiable GitHub username
cross_disciplinary_pairing.discipline	✅	The deep-tech discipline your model serves
cross_disciplinary_pairing.load_bearing	✅	true if the pairing is integral to the submission, not cosmetic
test_prompts	✅	Exactly 2 prompts in your chosen domain. Organizers will add 2 hidden prompts to test for overfitting.
model.runtime	✅	Must be llama.cpp. No other runtime is accepted.
model.quantization	✅	Must be a GGUF quantization format (e.g. GGUF Q4_K_M, GGUF Q5_K_M)
model.parameters_estimate	✅	Approximate parameter count (e.g. 135M, 1.1B, 7B)
model.packaging	✅	How the model is packaged. One of: docker_image, docker_build_from_repo, binary_bundle
_runtime.model_path	✅	Relative path from repo root to your .gguf file (e.g. model/my-model.gguf)

---

📥 download_model.sh

This script must download your model weight file to the model/ directory.

Rules:

• Must be idempotent — safe to run multiple times without re-downloading.
• Must work without any credentials — your weights must be publicly accessible.
• The downloaded file path must exactly match _runtime.model_path in metadata.json.

Recommended hosting options for your weights:

• Hugging Face — public model repos (free, best for GGUF files)
• GitHub Release Assets — attach the .gguf file to a GitHub Release
• Any stable public URL (GCS public bucket, S3 public object, etc.)

---

📄 REPORT.md

Your technical writeup. Judges and the LLM-based audit system will read this to understand your submission. Cover:

1. Problem — What problem are you solving? Who is the target user in an African context?
2. Design Decisions — What model did you start from? Why that quantization level? What alternatives did you evaluate?
3. Constraints — What hardware, connectivity, or data constraints shaped your approach?
4. Benchmarks — What inference speed and memory numbers did you observe on your development machine?

Keep it factual and specific. One to three pages is ideal.

---

🧪 Local Testing

The ADTC profiler is open source. Install it directly from the official repository:

pip install "git+https://github.com/Africa-Deep-Tech-Foundation/adtc-profiler.git"

Then run a local smoke test before submitting:

# 1. Download your weights
bash download_model.sh

# 2. Run the profiler in participant mode
adtc-profiler run \
  --submission . \
  --mode participant \
  --output submission.json \
  --skip-accuracy

# 3. Review your report
cat submission.json

A valid run produces a submission.json with "measured_on": "participant_laptop".

The profiler source code, including the thermal monitoring logic and scoring formulas, is publicly readable at: github.com/Africa-Deep-Tech-Foundation/adtc-profiler

---

⚠️ Rules

1. Public repository required. Your repository must be public at the time of evaluation.
2. No model weights in git. Add *.gguf and model/ to your .gitignore. The evaluator downloads weights fresh via download_model.sh.
3. 100% offline during evaluation. Your model must run with zero external network dependencies during our testing window. download_model.sh runs before the profiler starts, but once profiling begins, no outbound requests are permitted.
4. llama.cpp only. All models must use GGUF weights and run through llama.cpp. No other runtime is supported by our evaluation framework.
5. 8 GB RAM limit. Your model must run within the standard laptop profile (4 vCPU, 8 GB RAM, integrated GPU only). Out-of-memory errors during evaluation result in automatic disqualification.
6. No size restriction. There is no parameter count or file size cap — but the 8 GB RAM constraint is strict. Plan your quantization level accordingly.
7. Two test prompts required. Your metadata.json must include exactly 2 prompts in the test_prompts array. Organizers will generate 2 additional hidden prompts within your domain. All 4 are used for scoring.

---

🆘 Support

Open an issue in this repository or contact the ADTF team at challenge@africadeeptech.org.

View the full eligibility rules at adtc-2026.devpost.com/rules.

---

📄 License

This template is licensed under the terms of the GNU GPL v3 License.