Update README.md

README.md

@@ -10,6 +10,7 @@ tags:
 - slm
 - conversation
 - chat
+- gguf
 base_model:
 - dnotitia/Llama-DNA-1.0-8B-Instruct
 library_name: transformers
@@ -17,150 +18,52 @@ pipeline_tag: text-generation
 ---
 
 # DNA 1.0 8B Instruct
-<br>
+
 <p align="center">
 <img src="assets/dna-logo.png" width="400" style="margin: 40px auto;">
 </p>
-<br>
-
-## Introduction
-
-We introduce **DNA 1.0 8B Instruct**, a state-of-the-art (**SOTA**) bilingual language model optimized for both Korean and English languages, developed and released by **Dnotitia Inc.** This model is based on the Llama architecture and has been meticulously enhanced through various advanced training techniques to excel in language understanding and generation tasks.
-
-The DNA 1.0 8B Instruct model has undergone a sophisticated development process:
-
-- **Model Merging via SLERP:** Combined with Llama 3.1 8B Instruct using spherical linear interpolation to enhance performance.
-- **Knowledge Distillation (KD):** Utilizing Llama 3.1 405B as the teacher model to improve knowledge representation.
-- **Continual Pre-Training (CPT):** Trained on a high-quality Korean dataset to boost language capabilities.
-- **Supervised Fine-Tuning (SFT):** Aligned with human preferences through fine-tuning on curated data.
-- **Direct Preference Optimization (DPO):** Enhanced instruction-following abilities for better user interaction.
-
-Each model supports long-context processing of up to **131,072 tokens (128K)**, enabling it to handle extensive conversational histories and long documents effectively.
 
-<br>
-
-## Evaluation
-
-We evaluated DNA 1.0 8B Instruct against other prominent language models of similar sizes across various benchmarks, including Korean-specific tasks and general language understanding metrics.
-
-<br>
-
-<table>
-  <tr>
-    <th>Language</th>
-    <th>Benchmark</th>
-    <th>dnotitia<br>DNA 1.0<br>8B Instruct</th>
-    <th>EXAONE 3.5<br>7.8B</th>
-    <th>Qwen 2.5<br>7B</th>
-    <th>Llama 3.1<br>8B</th>
-    <th>Mistral<br>7B</th>
-  </tr>
-  <tr>
-    <td rowspan="5">Korean</td>
-    <td>KMMLU</td>
-    <td align="center"><strong>53.26</strong></td>
-    <td align="center">45.30</td>
-    <td align="center">45.66</td>
-    <td align="center">41.66</td>
-    <td align="center">31.45</td>
-  </tr>
-  <tr>
-    <td>KMMLU-Hard</td>
-    <td align="center"><strong>29.46</strong></td>
-    <td align="center">23.17</td>
-    <td align="center">24.78</td>
-    <td align="center">20.49</td>
-    <td align="center">17.86</td>
-  </tr>
-  <tr>
-    <td>KoBEST</td>
-    <td align="center"><strong>83.40</strong></td>
-    <td align="center">79.05</td>
-    <td align="center">78.51</td>
-    <td align="center">67.56</td>
-    <td align="center">63.77</td>
-  </tr>
-  <tr>
-    <td>Belebele</td>
-    <td align="center"><strong>57.99</strong></td>
-    <td align="center">40.97</td>
-    <td align="center">54.85</td>
-    <td align="center">54.70</td>
-    <td align="center">40.31</td>
-  </tr>
-  <tr>
-    <td>CSAT QA</td>
-    <td align="center">43.32</td>
-    <td align="center">40.11</td>
-    <td align="center"><strong>45.45</strong></td>
-    <td align="center">36.90</td>
-    <td align="center">27.27</td>
-  </tr>
-  <tr>
-    <td rowspan="3">English</td>
-    <td>MMLU</td>
-    <td align="center">66.64</td>
-    <td align="center">65.27</td>
-    <td align="center"><strong>74.26</strong></td>
-    <td align="center">68.26</td>
-    <td align="center">62.04</td>
-  </tr>
-  <tr>
-    <td>MMLU Pro</td>
-    <td align="center"><strong>43.05</strong></td>
-    <td align="center">40.73</td>
-    <td align="center">42.50</td>
-    <td align="center">40.92</td>
-    <td align="center">33.49</td>
-  </tr>
-  <tr>
-    <td>GSM8K</td>
-    <td align="center"><strong>80.52</strong></td>
-    <td align="center">65.96</td>
-    <td align="center">75.74</td>
-    <td align="center">75.82</td>
-    <td align="center">49.66</td>
-  </tr>
-</table>
-
-- The **highest scores** are in **bold**.
-
-<br>
+**DNA 1.0 8B Instruct** is a <u>state-of-the-art (**SOTA**)</u> bilingual language model based on Llama architecture, specifically optimized for Korean language understanding and generation, while also maintaining strong English capabilities. The model was developed through a sophisticated process involving model merging via spherical linear interpolation (**SLERP**) with Llama 3.1 8B Instruct, and underwent knowledge distillation (**KD**) using Llama 3.1 405B as the teacher model. It was extensively trained through continual pre-training (**CPT**) with a high-quality Korean dataset. The training pipeline was completed with supervised fine-tuning (**SFT**) and direct preference optimization (**DPO**) to align with human preferences and enhance instruction-following abilities.
 
-**Evaluation Protocol**
+DNA 1.0 8B Instruct was fine-tuned on approximately 10B tokens of carefully curated data and has undergone extensive instruction tuning to enhance its ability to follow complex instructions and engage in natural conversations.
 
-For easy reproduction of our evaluation results, we list the evaluation tools and settings used below:
+- **Developed by:** Dnotitia Inc.
+- **Supported Languages:** Korean, English
+- **Model Release Date:** Dec 10, 2024
+- **Vocab Size:** 128,256
+- **Context Length:** 131,072 tokens (128k)
+- **License:** CC BY-NC 4.0
 
-| Benchmark   | Evaluation Setting | Metric                              | Evaluation Tool    |
-|-------------|--------------------|-------------------------------------|--------------------|
-| KMMLU       | 5-shot             | `macro_avg` / `exact_match`         | `lm-eval-harness`  |
-| KMMLU-Hard  | 5-shot             | `macro_avg` / `exact_match`         | `lm-eval-harness`  |
-| KoBEST      | 5-shot             | `macro_avg` / `f1`                  | `lm-eval-harness`  |
-| Belebele    | 0-shot             | `accuracy`                          | `lm-eval-harness`  |
-| CSAT QA     | 0-shot             | `accuracy_normalized`               | `lm-eval-harness`  |
-| MMLU        | 5-shot             | `macro_avg` / `accuracy`            | `lm-eval-harness`  |
-| MMLU Pro    | 5-shot             | `macro_avg` / `exact_match`         | `lm-eval-harness`  |
-| GSM8K       | 5-shot             | `accuracy` / `exact_match`          | `lm-eval-harness`  |
-
-<br>
+<p align="center">
+<img src="assets/training-procedure.png" width="600" style="margin: 40px auto;">
+</p>
 
 ## Quickstart
 
 We offer weights in `F32`, `F16` formats and quantized weights in `Q8_0`, `Q6_K`, `Q5_K`, `Q4_K`, `Q3_K` and `Q2_K` formats.
 
-You can download the GGUF weights as follows:
+You can run GGUF weights with `llama.cpp` as follows:
+
+1. Install `llama.cpp`. Please refer to the [llama.cpp repository](https://github.com/ggerganov/llama.cpp) for more details.
+
+2. Download DNA 1.0 8B Instruct model in GGUF format.
 
 ```bash
 # Install huggingface_hub if not already installed
-pip install huggingface_hub
+$ pip install huggingface_hub[cli]
 
 # Download the GGUF weights
-huggingface-cli download dnotitia/Llama-DNA-1.0-8B-Instruct-GGUF \
-    --include "DNA-1.0-8B-Instruct-Q8_0.gguf" \
+$ huggingface-cli download dnotitia/Llama-DNA-1.0-8B-Instruct-GGUF \
+    --include "Llama-DNA-1.0-8B-Instruct-Q8_0.gguf" \
     --local-dir .
 ```
 
-<br>
+3. Run the model with `llama.cpp` in conversational mode.
+
+```bash
+$ llama-cli -cnv -m ./Llama-DNA-1.0-8B-Instruct-Q8_0.gguf \
+    -p "You are a helpful assistant, Dnotitia DNA."
+```
 
 ## Run Locally