arabic-semantic-chunker-qwen1.5b

Arabic Semantic Chunker — Qwen2.5-1.5B (LoRA Fine-tuned)

A production-grade, fine-tuned LLM based on Qwen/Qwen2.5-1.5B-Instruct, optimized specifically for the task of Semantic Chunking across Arabic, English, and mixed-language (code-switched) texts. It strictly guarantees structured, deterministic JSON outputs matching a predefined schema.

---

Table of Contents

• Model Overview
• Intended Use
• Training Details
• Dataset Architecture
• Infrastructure & Hardware
• Load Testing & Performance
• Usage Guide
• Output Schema
• Limitations & Boundary Conditions

---

Model Overview

This model acts as a dedicated LoRA Adapter layered over Qwen2.5-1.5B-Instruct. When fed a text corpus, it splits it into logically self-contained, semantically coherent chunks. It preserves 100% of the input text—no paraphrasing, translation, summarization, or omission occurs.

---

Intended Use

This model is engineered to serve as a core component in enterprise-grade Arabic/multilingual NLP pipelines, specifically:

• Retrieval-Augmented Generation (RAG): Chunking unstructured documents into highly cohesive text segments prior to vector embedding generation.
• Advanced Preprocessing: A smart, context-aware alternative to naive rule-based text splitters (e.g., regex, token count limits) which frequently break up semantic contexts in complex Arabic phrasing.
• Technical Document Analysis: Flawlessly handling technical Arabic material heavily interspersed with English terminology.
• Information Extraction (IE): Isolating independent factual claims or narrative transitions for downstream analytical tasks.

Caution: This model is strictly a sequence segmenter. It is not designed for abstractive generation, translation, open-ended QA, or text summarization.

---

Training Details

LoRA Configuration

Supervised Fine-Tuning (SFT) Hyperparameters

The final adapter utilized in production corresponds to checkpoint-1000, chosen for its optimal evaluation convergence and stable validation loss profile without indicating signs of overfitting.

---

Dataset Architecture

The dataset was constructed via a rigorous Knowledge Distillation pipeline utilizing gemini-3.1-flash-lite as the Teacher Model to generate high-quality ground-truth mappings (Raw Text → Syntactically Valid JSON Chunks).

Data Sources & Distribution

Three distinct text styles were synthesized to ensure cross-domain generalization:

Data Splits

Each training sequence wraps the original raw layout into an instruction text prompt, with a strictly deterministic, character-exact JSON schema mapping as the targeted output.

---

Infrastructure and Hardware

Integrating Unsloth was vital to successfully execute this training pipeline on a commodity T4 GPU's restricted VRAM profile (15GB), providing nearly a 2x acceleration advantage alongside minimal hardware activation overhead.

---

Load Testing & Performance

Following convergence, the model was deployed via the high-concurrency vLLM engine and stress-tested using Locust to simulate realistic concurrent production workloads.

Serving Configuration (vLLM)

Performance Benchmark Under Stress

A sustained load test simulating simultaneous clients generating continuous validation payloads yields the following verified metrics:

Architectural Note: These benchmarks were successfully conducted under a concurrent load of 20 users. vLLM's Continuous Batching and PagedAttention maximized serving stability, keeping the GPU KV Cache usage at a highly efficient ~3.2% with zero out-of-memory (OOM) or timeout constraints.

---

Usage Guide

Loading via Unsloth (Recommended for Inference)

from unsloth import FastLanguageModel

# Load the model and tokenizer natively using Unsloth
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "Mo-Abdelfattah/arabic-semantic-chunker-qwen1.5b",
    max_seq_length = 3500,
    dtype          = None,
    load_in_4bit   = False,
)
FastLanguageModel.for_inference(model)

Loading via Standard Hugging Face Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model first
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

# Wrap with the LoRA adapter weights
model = PeftModel.from_pretrained(base_model, "Mo-Abdelfattah/arabic-semantic-chunker-qwen1.5b")
tokenizer = AutoTokenizer.from_pretrained("Mo-Abdelfattah/arabic-semantic-chunker-qwen1.5b")

Production Serving via vLLM CLI

vllm serve "Qwen/Qwen2.5-1.5B-Instruct" \
  --dtype=half \
  --gpu-memory-utilization 0.8 \
  --max-model-len 2048 \
  --max-lora-rank 64 \
  --enable-lora \
  --lora-modules arabic-chunker="Mo-Abdelfattah/arabic-semantic-chunker-qwen1.5b"

---

Output Format

The model strictly adheres to outputting valid, parsable JSON matching the following structural schema:

{
  "original_text_length": 2345,
  "semantic_chunks": [
    "First coherent semantic segment extracted from the text...",
    "Second logical chunk containing continuing semantic context..."
  ]
}

---

Limitations & Boundary Conditions

• Strict Task Bound: The model is exclusively trained for text sequence segmentation. Abstractive reasoning, creative generation, or open-domain synthesis tasks are explicitly unsupported.
• Context Length Restrictions: The training pipeline sliced input sources at a maximum boundary of 1,500 characters per document sample; performance on exceedingly long inputs (e.g., >10k tokens) may degrade.
• Distillation Dependency: The training data utilizes synthesized extractions from gemini-3.1-flash-lite, meaning that any localized logical parsing bias present in the teacher model may occasionally be inherited.

---

Author

Mohamed Abdelfattah AI & Data Platform Engineer specializing in high-throughput Arabic NLP architectures, robust LLM alignment pipelines, and optimal distributed serving infrastructures.