Code for the ArabicNLP 2026 paper rabia_naz at StanceEval-2026: An Ensemble of Fine-Tuned Open LLMs for Arabic Stance Detection on Seen Targets.
The system scored 0.7564 Favg2 on StanceEval-2026 Track 1, where the test target is related to a training target but absent from the training data. The prediction is a vote over three QLoRA fine-tuned models, Qwen2.5-7B, ALLaM-7B, and a sarcasm-aware Qwen2.5-7B, adapted to the test target with pseudo-labels accepted only where independent models agree.
| File | Purpose |
|---|---|
01_qwen_selftrained.py | Trains the self-trained Qwen2.5-7B |
02_allam.py | Trains ALLaM-7B |
03_sarcasm_and_vote.py | Trains the sarcasm-aware model, builds the final vote and submission file |
Each script runs on a single Kaggle T4 in three to four hours and installs its own dependencies.
| Parameter | Value |
|---|---|
| Adapter | LoRA, r=16, alpha=32 |
| Quantization | 4-bit NF4 |
| Learning rate | 2e-4 |
| Effective batch size | 16 |
| Seed | 42 |
| Base models | Qwen/Qwen2.5-7B-Instruct, ALLaM-AI/ALLaM-7B-Instruct-preview |
Code for the ArabicNLP 2026 paper rabia_naz at StanceEval-2026: An Ensemble of Fine-Tuned Open LLMs for Arabic Stance Detection on Seen Targets.
The system scored 0.7564 Favg2 on StanceEval-2026 Track 1, where the test target is related to a training target but absent from the training data. The prediction is a vote over three QLoRA fine-tuned models, Qwen2.5-7B, ALLaM-7B, and a sarcasm-aware Qwen2.5-7B, adapted to the test target with pseudo-labels accepted only where independent models agree.
| File | Purpose |
|---|---|
01_qwen_selftrained.py | Trains the self-trained Qwen2.5-7B |
02_allam.py | Trains ALLaM-7B |
03_sarcasm_and_vote.py | Trains the sarcasm-aware model, builds the final vote and submission file |
Each script runs on a single Kaggle T4 in three to four hours and installs its own dependencies.
| Parameter | Value |
|---|---|
| Adapter | LoRA, r=16, alpha=32 |
| Quantization | 4-bit NF4 |
| Learning rate | 2e-4 |
| Effective batch size | 16 |
| Seed | 42 |
| Base models | Qwen/Qwen2.5-7B-Instruct, ALLaM-AI/ALLaM-7B-Instruct-preview |
