---
language:
- ar
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
library_name: transformers
tags:
- fine-tuned
- arabic
- financial-nlp
- information-extraction
- lora
- llama-factory
datasets:
- custom
pipeline_tag: text-generation
---

# qwen2.5-arabic-finance-news-parser

A fine-tuned version of [Qwen/Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) for **structured information extraction from Arabic financial news articles**.

## Model Description

This model was fine-tuned using **LoRA** via [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) on a dataset of ~2,792 Egyptian stock-market news articles. Given a news article and a JSON output schema, the model extracts structured data such as company name, event type, sentiment, financial figures, and a short summary.

## Training Details

| Setting | Value |
|---|---|
| Base model | Qwen/Qwen2.5-1.5B-Instruct |
| Fine-tuning method | LoRA (rank 64, all targets) |
| Dataset size | 2,792 samples (2,700 train / 92 val) |
| Epochs | 3 |
| Learning rate | 1e-4 (cosine scheduler) |
| Max sequence length | 3,500 tokens |
| Hardware | Kaggle T4 GPU |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch, json

model_id = "tegana/qwen2.5-arabic-finance-news-parser"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", torch_dtype=torch.bfloat16
)

article = "القاهرة - واصل جهاز مستقبل مصر للتنمية المستدامة..."

output_scheme = json.dumps({
    "company_name": "اسم الشركة",
    "event_type": "acquisition|earnings|dividends|...",
    "sentiment": "positive|negative|neutral",
    "impact_level": "high|medium|low",
    "short_summary": "ملخص من 3 إلى 5 جمل"
}, ensure_ascii=False)

messages = [
    {"role": "system", "content": (
        "You are a professional Arabic financial news parser.\n"
        "Extract structured information and return ONLY a valid JSON object."
    )},
    {"role": "user", "content": f"## Article:\n{article}\n\n## Output Scheme:\n{output_scheme}\n\n## Output JSON:"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(inputs.input_ids, max_new_tokens=512, do_sample=False)

out_ids = [o[len(i):] for i, o in zip(inputs.input_ids, out)]
print(tokenizer.batch_decode(out_ids, skip_special_tokens=True)[0])
```

## Supported Event Types

`earnings` · `capital_increase` · `capital_decrease` · `dividends` · `acquisition` · `sale_of_stake` · `financing` · `project` · `board_decision` · `regulatory_approval` · `analysis_financial` · `stock_exchange_decision` · `other`

## Limitations

- Trained primarily on Egyptian stock-market news; may underperform on other Arabic financial dialects.
- Numerical extraction quality depends on how clearly figures appear in the source text.

## License

Apache 2.0 — same as the base model.