SUHAIL-14B-preview

14B Arabic LLM – LoRA fine-tuned from Qwen-3-14B-Base for instruction following and human-preference alignment

---

TL;DR

• Base model: Qwen-3-14B-Base (Transformer decoder, Rotary Positional Embeddings)
• Fine-tuning: Two-stage Low-Rank Adaptation (LoRA)
  1. Supervised Fine-Tuning (SFT) on a curated Arabic/English instruction dataset
  2. Human Preference Alignment using binary accept/reject feedback
• Data selection: Employed a state-of-the-art encoder-based reranker to filter the Efficient Instruction-Tuning corpus via Style-Aligned Response Ranking, retaining only stylistically coherent, high-quality samples
• Context window: 32k tokens
• License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
• Intended use: Arabic content generation, multi-turn tool use (Agentic System), conversational agents, educational tools, and research (non-commercial only)
• Training samples: 33k (SFT), 66k (human preference alignment)
• Training cost: Less than $500

---

Table of Contents

1. Model Description
2. Quick Start
3. Limitations & Biases
4. License
5. Citation
6. Changelog

---

Table of contents

1. Model description
2. Quick start
3. Limitations & biases
4. License
5. Citation
6. Changelog

---

Model Description

SUHAIL-14B-preview extends the open-weight Qwen-3-14B-Base to better support Arabic instruction-following using Low-Rank Adaptation (LoRA). LoRA introduces small trainable matrices to linear layers as well as attention layers, keeping base weights frozen—enabling compact, efficient fine-tuning.

1 · Supervised Fine-Tuning (SFT)

We first conducted SFT on a high-quality instruction dataset in Arabic and English. This dataset was curated using Style-Aligned Response Ranking, a RoBERTa-based reranker that filters out stylistically incoherent or low-quality samples from the Instruction-Tuning corpus. This step enhanced factuality and stylistic consistency.

Result: Up to 22% performance improvements observed on internal benchmarks (e.g., IFEVAL).

2 · Human Preference Alignment

To align model behavior with user intent, we applied preference optimization using binary accept/reject feedback. This direct signal training guides the model toward generating helpful, honest, and harmless outputs, at low alignment cost.

3 · Integration of Reinforcement Learning with Verifiable Rewards (RLVR) for Large Language Models in Verifiable and Auditable Environments (TO-DO)

4 · Benchmarks (TO-DO)

Explicit benchmark scores are not yet included. We encourage users to evaluate the model in their specific contexts.

---

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

device = "cuda:0"
model_id = "01-ZeroOne/SUHAIL-14B-preview"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "اكتب ملخصًا بسيطًا عن الإنترنت باللغة العربية."
inputs = tokenizer(prompt, return_tensors="pt").to(device)
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

The LoRA adapters are merged into the checkpoint on the Hub for ease of use.

---

Limitations & biases

• Factual reliability – hallucinations remain. Verify critical information.
• Dialect coverage – best on Gulf & Egyptian Arabic; less data for Maghrebi and Levantine.
• Code completeness – suitable for small code snippets, but not guaranteed bug-free.
• Agentic Function Calling Coverage – Preliminary support included in SFT. Future updates aim to enhance reasoning and structured API calling capabilities.

---

License

Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) — non-commercial use only.

---

Citation

@software{Suhail2025,
  author = {ZeroOne AI},
  title  = {SUHAIL-14B-preview},
  year   = {2025},
  url    = {https://huggingface.co/01-ZeroOne/SUHAIL-14B-preview}
}

---

Changelog

Version	Date	Notes
v0.1	2025-07-05	Initial public LoRA-merged release (SFT + human-preference alignment; data filtered with Style-Aligned Response Ranking)

---

Maintained by Mohammed Almaghrabi, Founder of ZeroOne AI. This work was supported by Khalid Alharbi — contributions are welcome! To contribute, please email: [email protected]