DRU-RE-qwen35-9B

QLoRA adapter for Arabic relation extraction using Qwen/Qwen3.5-9B as the base model.

This repository stores the adapter, training/evaluation artifacts, prompt/GEPA experiments, ontology resources, and the code used for training and evaluation.

Best Checkpoint

The main adapter files at the repository root are from checkpoint-2000, which had the best full validation score.

checkpoint	validation rows	correct	micro F1	macro-style F1 field
step 1000	1687	1034	0.612922	0.503539
step 2000	1687	1177	0.697688	0.715525
step 2943	1687	1171	0.694132	0.709890

Repository Contents

• Root adapter files: best QLoRA adapter from checkpoint-2000.
• checkpoints/: saved checkpoints for step 1000, 2000, and 2943, including trainer state and optimizer state for resume.
• evals/: evaluation metrics for each saved step.
• predictions.txt: validation predictions from checkpoint 2000, formatted as triple_id<TAB>predicted_label.
• ontology/: wojood_ontology.json, relation_mapper.json, and relation_disambiguation.json.
• code/: training, evaluation, prompt optimization, metrics, and utility scripts.
• gepa/: GEPA prompt optimization summaries, best candidates, split summaries, and metrics.
• baseline_evals/: base-prompt evaluation metrics before finetuning.

Prediction File Format

predictions.txt contains one validation prediction per line:

11352	Location.located_in
11353	Location.located_in
4838	Location.located_in

Training Command

cd /root/workspace
python RE/finetune_qwen.py --config RE/finetune_config.yaml

To resume:

cd /root/workspace
python RE/finetune_qwen.py --config RE/finetune_config.yaml --resume-from-checkpoint latest

Inference Notes

Load this repository as a PEFT adapter on top of Qwen/Qwen3.5-9B. The prompt format and ontology filtering logic are included in code/prompt_baseline.py, code/ontology.py, and code/finetune_config.yaml.

The model should output exactly one relation label from the configured label set.

Known Weaknesses

The step-2000 evaluation still showed confusion around:

• PartOf.geopolitical_division versus no_relation.
• Location.headquartered_in versus Location.located_in.
• Overprediction of Personal.has_occupation for some no_relation examples.
• Low recall for several rare relation classes.

Important

This repo contains a QLoRA/PEFT adapter, not the full Qwen base model weights.