RightNow-Arabic-0.5B-Turbo/README.md

---
license: apache-2.0
language:
- ar
- en
base_model: Qwen/Qwen2.5-0.5B
pipeline_tag: text-generation
tags:
- arabic
- edge
- small-language-model
- sft
- dpo
- gguf
- qwen2
library_name: transformers
model-index:
- name: RightNow-Arabic-0.5B-Turbo
  results:
  - task:
      type: text-generation
      name: COPA-ar
    dataset:
      name: copa_ar
      type: copa_ar
    metrics:
    - type: acc_norm
      value: 58.4
      name: Accuracy (norm)
  - task:
      type: text-generation
      name: Arabic HellaSwag
    dataset:
      name: arabic_mt_hellaswag
      type: arabic_mt_hellaswag
    metrics:
    - type: acc_norm
      value: 26.0
      name: Accuracy (norm)
  - task:
      type: text-generation
      name: ArabicMMLU
    dataset:
      name: arabic_leaderboard_arabic_mmlu
      type: arabic_leaderboard_arabic_mmlu
    metrics:
    - type: acc
      value: 23.2
      name: Accuracy
---

<div align="center">

# RightNow-Arabic-0.5B-Turbo

### The smallest open Arabic-specialized decoder LLM

**518M parameters | 398 MB on disk (q4_k_m) | 635 tok/s on H100**

*Built by [RightNow AI](https://rightnowai.co)*

<img src="benchmark_pareto.png" width="700">

</div>

---

## What is this?

RightNow-Arabic-0.5B-Turbo is a **518M-parameter Arabic-specialized language model** built on top of Qwen2.5-0.5B via vocabulary injection, continued pretraining, supervised fine-tuning, and direct preference optimization. It is the **smallest open Arabic-specialized decoder LLM** with publicly available weights.

The model targets **edge deployment**: phones, laptops, embedded devices, and browsers. Quantized to q4_k_m it fits in 398 MB and generates 635 tokens/s at batch size 1.

## Key Features

- **27,032 new Arabic tokens** added via mean-subtoken initialization, cutting Arabic tokenizer fertility by 17.3% (2.18 to 1.80 tokens/word)
- **504M Arabic pretraining tokens** (Arabic Wikipedia) on 8xH100 SXM5 with FSDP + FlashAttention varlen + Liger fused kernels
- **129,116 Arabic instruction pairs** for SFT with response-only loss masking
- **6,750 Arabic preference pairs** for DPO
- **Weight soup merging** (DPO 50%, SFT 25%, Pretrain 25%) for optimal accuracy
- **4 GGUF quantizations** for instant edge deployment

## Benchmarks

Evaluated with [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) v0.4.11, `apply_chat_template=True`, `limit=200`, `acc_norm` preferred.

### Head-to-head comparison

| Model | Params | COPA-ar | HellaSwag-ar | ArabicMMLU | Mean |
|-------|--------|---------|-------------|------------|------|
| **RightNow-Arabic-0.5B-Turbo (ours)** | **518M** | **58.4%** | **26.0%** | 23.2% | **35.9%** |
| Qwen2.5-0.5B-Instruct | 494M | 53.9% | 22.5% | **26.0%** | 34.1% |
| Falcon-H1-0.5B-Instruct | 524M | 44.9% | 23.0% | 24.2% | 30.7% |
| Falcon-H1-1.5B-Instruct | 1.5B | 58.4% | 27.5% | 32.7% | 39.5% |
| AceGPT-7B-chat | 7B | 69.7% | 27.0% | 35.0% | 43.9% |
| ALLaM-7B-Instruct | 7B | 68.5% | 29.0% | 52.2% | 49.9% |
| SILMA-9B-Instruct | 9B | 69.7% | 38.0% | 52.9% | 53.5% |

**Among 0.5B models**: best on COPA-ar (+4.5 vs Qwen), best on HellaSwag-ar (+3.5 vs Qwen), best mean (+1.8 vs Qwen, +5.2 vs Falcon).

**Ties Falcon-H1-1.5B** on COPA-ar (both 58.4%) at one-third the parameters.

**Recovers 67%** of SILMA-9B mean accuracy at 5.8% of the parameters.

## Available Formats

| Format | Size | Speed (tok/s, bs=1, H100) | Use case |
|--------|------|---------------------------|----------|
| bf16 | 1.04 GB | 82 (HF generate) | Fine-tuning, research |
| int8 | 664 MB | -- | Reduced memory inference |
| GGUF f16 | 988 MB | 582 | Maximum quality |
| **GGUF q8_0** | **525 MB** | **646** | Best speed |
| GGUF q5_k_m | 419 MB | 634 | Balanced |
| **GGUF q4_k_m** | **398 MB** | **635** | Smallest footprint |

## Quick Start

### With Transformers (bf16)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "RightNowAI/RightNow-Arabic-0.5B-Turbo"
tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="bf16")
model = AutoModelForCausalLM.from_pretrained(
    model_id, subfolder="bf16",
    torch_dtype="bfloat16", device_map="auto"
)

messages = [
    {"role": "system", "content": "أنت مساعد ذكي يجيب باللغة العربية الفصحى"},
    {"role": "user", "content": "ما هي عاصمة المملكة العربية السعودية؟"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

### With llama.cpp (GGUF)

```bash
# Download the q8_0 quantization (best speed)
huggingface-cli download RightNowAI/RightNow-Arabic-0.5B-Turbo \
  gguf/RightNow-Arabic-0.5B-Turbo-q8_0.gguf --local-dir .

# Run inference
./llama-cli -m RightNow-Arabic-0.5B-Turbo-q8_0.gguf \
  -p "ما هي أكبر مدينة في مصر؟" \
  -n 128 --temp 0.7
```

## Training Pipeline

```
Qwen2.5-0.5B (494M, 151,665 vocab)
        |
        v
Tokenizer Surgery (+27,032 Arabic tokens -> 178,697 vocab)
  - SentencePiece unigram 32k on 12.5 GB Arabic corpus
  - Mean-subtoken embedding initialization
  - Fertility: 2.18 -> 1.80 tok/word (-17.3%)
        |
        v
Continued Pretraining (504M arwiki tokens)
  - 2,500 steps, 8xH100 SXM5
  - FSDP _HYBRID_SHARD_ZERO2 + FlashAttention varlen + Liger
  - Loss: 14.21 -> 1.69 | Wall time: 6h 57m
        |
        v
Supervised Fine-Tuning (129,116 instructions)
  - 5 datasets: evol-instruct-arabic, alpaca-gpt4-arabic,
    sharegpt-arabic, CIDAR, aya_dataset
  - Response-only loss masking (72.1% of tokens carry loss)
  - 5 epochs, 418 steps | Wall time: 12m
        |
        v
Direct Preference Optimization (6,750 pairs)
  - argilla-dpo-mix-7k-arabic
  - 2 epochs, 844 steps | Wall time: 34m
        |
        v
Weight Soup Merging
  - Linear(DPO 0.5, SFT 0.25, Pretrain 0.25)
  - +0.44 points mean accuracy over DPO alone
        |
        v
Export: bf16, int8, GGUF {f16, q8_0, q5_k_m, q4_k_m}
```

## Training Data

| Dataset | Examples/Tokens | Use |
|---------|----------------|-----|
| Arabic Wikipedia (wikimedia/wikipedia 20231101.ar) | 504M tokens | Continued pretraining |
| FreedomIntelligence/evol-instruct-arabic | 59,022 | SFT |
| FreedomIntelligence/alpaca-gpt4-arabic | 49,969 | SFT |
| FreedomIntelligence/sharegpt-arabic | 5,231 | SFT |
| arbml/CIDAR | 10,000 | SFT |
| CohereForAI/aya_dataset (Arabic) | 4,947 | SFT |
| 2A2I/argilla-dpo-mix-7k-arabic | 6,750 | DPO |

## Limitations

- **Knowledge ceiling**: At 518M parameters, ArabicMMLU-style knowledge tasks lag 7B+ models by 12-30 points. This is a parameter-count limit, not a training limit.
- **MSA only**: Trained on Wikipedia (Modern Standard Arabic). Dialects (Egyptian, Gulf, Levantine) get MSA responses.
- **504M pretraining tokens**: Below Chinchilla-optimal ratio. More Arabic data would improve knowledge tasks.
- **DPO was weak**: 6,750 machine-translated preference pairs provided minimal signal at 0.5B scale. The weight soup merge was more impactful.
- **GGUF tile alignment**: q4_k_m and q5_k_m fall back to higher-bit quantization for 144/290 tensors due to the expanded vocabulary not aligning with k-quant tile sizes.

## Hardware

All training ran on a single Nebius `gpu-h100-sxm` node:
- 8x NVIDIA H100 80 GB SXM5 HBM3, NVLink4
- 128 vCPUs, 1.5 TiB RAM
- CUDA 13.0, PyTorch 2.11, flash-attn 2.8.3, transformers 5.5.0

## Citation

```bibtex
@article{jaber2025rightnow,
  title={RightNow-Arabic-0.5B-Turbo: An Open Sub-1B Arabic Language Model via Vocabulary Injection and Edge-First Deployment},
  author={Jaber, Jaber and Jaber, Osama},
  year={2025},
  url={https://huggingface.co/RightNowAI/RightNow-Arabic-0.5B-Turbo}
}
```

## License

Apache 2.0 (same as the base Qwen2.5-0.5B model).

---

*Built by [RightNow AI](https://rightnowai.co)*