Upload README.md with huggingface_hub

README.md

@@ -7,30 +7,50 @@ tags:
 - mmlu
 - exaone
 - amd-mi325
-pretty_name: EXAONE 1.2B Quantized MMLU Results
+- gptq
+pretty_name: EXAONE 4.0 1.2B Quantized MMLU Evaluation
 size_categories:
 - n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: mmlu_refined.csv
 ---
 
-# EXAONE-4.0-1.2B-GPTQ MMLU Evaluation
+# 🚀 EXAONE-4.0-1.2B-Quantized-MMLU Evaluation Results
 
-This repository contains refined evaluation results for the **EXAONE-4.0-1.2B-GPTQ-W8A16** model.
+This repository contains the refined evaluation results for the **EXAONE-4.0-1.2B-GPTQ-W8A16** model using the MMLU benchmark.
 
-## 🚀 Evaluation Setup
-- **Benchmark:** MMLU (Massive Multitask Language Understanding)
-- **Prompting:** 5-shot
-- **Hardware:** AMD Instinct MI325 OAM
+## 📊 Overview
+The evaluation was conducted to measure the model's multitask language understanding capabilities across 57 different subjects. To ensure stability and formatting consistency, we used a **5-shot** prompting approach.
+
+### 💻 Hardware & Software
+- **Model:** [MangoLab/EXAONE-4.0-1.2B-GPTQ-W8A16](https://huggingface.co/MangoLab/EXAONE-4.0-1.2B-GPTQ-W8A16)
+- **Accelerator:** AMD Instinct MI325 OAM
 - **Quantization:** GPTQ (W8A16 / fp8_w8a8)
+- **Framework:** `lm-evaluation-harness`
 
-## 📊 Quick Results
-The full breakdown of 57 subjects can be found in `mmlu_refined.csv`.
+## 📈 Performance Summary
+The table below shows a summary of the evaluation environment. Detailed per-subject accuracy can be viewed in the **Dataset Viewer** above.
 
-| Metric | Value |
+| Category | Details |
 | :--- | :--- |
-| **Tool** | lm-evaluation-harness |
-| **Accelerator** | AMD MI325 |
-| **Status** | Completed |
+| **Benchmark** | MMLU (Massive Multitask Language Understanding) |
+| **Prompting** | 5-shot |
+| **Dtype** | FP8 / W8A16 |
+| **Evaluation Date** | 2026-01-24 |
+
+## 📂 Directory Structure
+- `mmlu_refined.csv`: The main result file containing subject-wise accuracy (%).
+- `raw_data/`: Original JSON output files from the evaluation process for reproducibility.
 
-## 📂 File Structure
-- `mmlu_refined.csv`: Key performance metrics per subject (Accuracy %).
-- `raw_data/`: Original JSON output from the evaluation harness.
+---
+## How to reproduce
+To run the same evaluation, use the following command with `lm-evaluation-harness`:
+```bash
+accelerate launch -m lm_eval --model hf \
+    --model_args pretrained=MangoLab/EXAONE-4.0-1.2B-GPTQ-W8A16,trust_remote_code=True \
+    --tasks mmlu \
+    --num_fewshot 5 \
+    --batch_size auto