README.md

---
license: apache-2.0
language:
- km
pipeline_tag: text-generation
authors:
- sovannpanhaseng
datasets:
- attentionlab/fineweb-2-khmer-extended
---

# Bayon (បាយ័ន)

**Bayon** is the first truly native Khmer language model. Unlike many existing models that are simply fine-tuned versions of Western or multilingual architectures, this model was built and pretrained entirely from scratch by Cambodians, for Cambodia.

---

## 🚀 Key Features

* **100% Pretrained From Scratch:** This model was not fine-tuned on top of an existing English-centric LLM. It was trained from the ground up natively on Khmer text data, giving it a deep, foundational grasp of the language's syntax and nuances.
* **Custom 5K Khmer Tokenizer:** Built with a custom-engineered 5,000-token vocabulary optimized specifically for the Khmer script. This eliminates the "token tax" commonly found in Western tokenizers, making inference significantly faster and more efficient for Khmer text.
* **Basic Generative Capability:** Ready out of the box for basic text completion tasks in Khmer.

---

## 📚 Training Data & Dataset Curation

This model was pretrained on a highly curated corpus consisting of **1.361 billion tokens** (as processed by the model's custom tokenizer).

To ensure high-quality generation and robust linguistic understanding, the dataset was built using:

* **FineWeb-2 (Khmer Extension):** Leveraging the massive, cleaned web-scale data from the FineWeb-2 initiative as our core foundation.
* **Custom Cleanup Pipelines:** We applied rigorous, proprietary filtering and deduplication methods tailored specifically to the Khmer script. This process stripped out machine-translated gibberish, HTML noise, and non-Khmer text, leaving a pristine dataset representing authentic language use.

---

## 🏗️ Model Architecture

The model is built on a highly optimized, deep-yet-narrow **Gemma-style** autoregressive decoder architecture. While many lightweight models sacrifice depth to reduce parameter counts, this architecture prioritizes depth (28 layers) to capture complex, long-range structural dependencies unique to the Khmer language, while maintaining a lean hidden dimension to stay incredibly fast and memory-efficient.

### Key Architectural Features

* **Grouped Query Attention (GQA):** Features 8 attention heads for queries but scales down to 2 heads for Keys and Values (KV). This significantly cuts down the KV-cache memory footprint during generation, allowing for faster inference and larger batch sizes.
* **GeGLU Activation:** The feed-forward network (MLP) utilizes Gated Linear Units with GELU activation functions (`gate_proj` paired with `up_proj` before projecting down), which has been shown to offer superior semantic representation over standard ReLU or vanilla GELU.
* **Rotary Position Embeddings (RoPE):** Implements dynamic rotary embeddings to inject positional context directly into the attention mechanism, supporting a context window of up to 2,048 tokens.
* **Root Mean Square Normalization (RMSNorm):** Applied at both the input and post-attention stages of every decoder layer to stabilize gradient flows and speed up training convergence without the computational overhead of standard LayerNorm.

---

### 📊 Hyperparameters at a Glance

| Hyperparameter | Value | Description |
| --- | --- | --- |
| **Parameters** | ~100M | Total trainable parameter count |
| **Layers (`num_hidden_layers`)** | 28 | Deep transformer stack for complex linguistic hierarchy |
| **Hidden Size (`hidden_size`)** | 512 | Width of the embedding and hidden states |
| **Intermediate Size** | 2,048 | Dimension of the GeGLU feed-forward layer |
| **Attention Heads ($Q$)** | 8 | Number of query heads |
| **Key-Value Heads ($K, V$)** | 2 | Grouped Query Attention (GQA) configuration |
| **Head Dimension** | 64 | Dimension per attention head |
| **Context Length (`max_position_embeddings`)** | 2,048 | Maximum sequence token window |
| **Vocabulary Size** | 5,000 | Custom localized Khmer-optimized vocabulary |
| **Rope Theta** | 10,000.0 | Base frequency for rotary position embeddings |

---

## 🛠️ How to Use

You can easily use this model utilizing the Hugging Face `transformers` pipeline ecosystem.

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

# 1. Specify the model repository path
model_id = "attentionlab/bayon"

# 2. Load the custom tokenizer and optimized model weights
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)

# 3. Format an example prompt (Basic QA / Text Generation)
prompt = "សួស្តី តើអ្នកអាច"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# 4. Generate sequences natively
with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=50,
        do_sample=True,
        temperature=0.2,
        top_k=20
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

```

---

## 🤝 Acknowledgments & Authors

This model is a proud step forward for the Cambodian AI ecosystem, developed independently by **[Seng Sovannpanha](https://huggingface.co/sovannpanhaseng)** to push the boundaries of Khmer Natural Language Processing (NLP).