README.md

---
license: other
library_name: custom
tags:
  - zindi
  - telco
  - track-a
  - agentic-workflow
  - competition-submission
  - lightgbm
  - qwen
---

# Track A Submission

## Environment Setup

Install the Python dependencies from this directory:

```bash
pip install -r requirements.txt
```

The code expects the organizer-provided Track A tool server at:

```text
https://localhost:8081/no
```

To override it, set `TRACK_A_SERVER_URL` or pass `--server_url`.

## Model Deployment

The Qwen3.5-35B-A3B base model is not included in this package. Deploy the local base model with vLLM using:

```bash
bash models/deploy.sh
```

Set `BASE_MODEL_PATH` before running the script if the model is not located at `/models/Qwen3.5-35B-A3B`.

The auxiliary Track A model bundle is stored at:

```text
models/model_v4_bundle.pkl
```

## Reproducing The Trained Model

The Phase 1 labelled training data is included at:

```text
data/Phase_1/train.json
```

To retrain the auxiliary Track A model bundle from scratch, run:

```bash
python train.py \
  --train_path data/Phase_1/train.json \
  --out models/model_v4_bundle.pkl \
  --experiment_name lgbm_v4 \
  --n_jobs -1
```

This trains the template classifier and candidate selector, writes experiment artifacts under `results/experiments/`, and places the final model bundle at:

```text
models/model_v4_bundle.pkl
```

## How To Run

Run the solution with the private Track A test file:

```bash
python run.py --input /path/to/test.json --output result
```

Optional useful arguments:

```bash
python run.py \
  --input /path/to/test.json \
  --output result \
  --server_url https://localhost:8081/no \
  --model_url http://localhost:8001/v1 \
  --model_name Qwen3.5-35B-A3B
```

## Expected Output

The runner writes:

```text
result/
  traces.json
  results.csv
  runtime.json
```

`results.csv` contains:

```csv
scenario_id,prediction
```

`runtime.json` is derived from the per-scenario execution timings recorded by the inference code.