---
license: other
license_name: aidctech
license_link: LICENSE
tags:
- unsloth
datasets:
- Phonepadith/laos-long-content
language:
- lo
base_model:
- google/gemma-3-12b-it
new_version: Phonepadith/aidc-llm-laos-10k-gemma-3-12b-it-v2
pipeline_tag: summarization
---

---
# 🧠 Lao Summarization Model  ສົນທະນາ - ສະຫລຸບເນື້ອຫາສຳລັບພາສາລາວ - Fine-tuned Gemma 3 12B IT (24,000 Pairs, Laos Content Input-Output)

This is a **Lao language summarization model** fine-tuned on the [`Phonepadith/laos_word_dataset`](https://huggingface.co/datasets/Phonepadith/laos_word_dataset), using the base model [`google/gemma-3-12b-it`](https://huggingface.co/google/gemma-3-12b-it). The model is designed to generate concise summaries from Lao language text.
**Scope**:
- 📚 ສະຫລຸບຂ່າວ
- 📚 ສະຫລຸບເອກະສານພາກລັດ
- 📚 ສະຫລຸບກອງປະຊຸມ
---

# 🧠 Lao AIDC-24K Fine-tuned Gemma-3-12B-IT
**Model ID**: `Phonepadith/aidc-llm-laos-24k-gemma-3-12b-it`  
**Base Model**: [`google/gemma-3b-it`](https://huggingface.co/google/gemma-3b-it)  
**Fine-tuned By**: [Phonepadith Phoummavong](https://huggingface.co/Phonepadith)

---

## 📌 Model Description

This model is a fine-tuned version of **Gemma-3-12B-IT-**, specifically adapted to understand and generate responses in **Lao language** 🇱🇦. It was trained using a curated dataset of over **5,000 high-quality Lao input-output pairs**, primarily focused on **AIDC (Artificial Intelligence and Digital Community)** topics.

**Key Features:**
- 🗣️ Fine-tuned for Lao language generation
- 📚 Suitable for summarization, question answering, general chat
- 🧠 Based on Google's powerful Gemma 3-12B Instruct model

---

## 🧾 Training Details

| Detail               | Value                            |
|----------------------|-----------------------------------|
| Base Model           | Gemma 3-12B Instruct              |
| Fine-tuning Method   | LoRA with PEFT (Unsloth)          |
| Dataset              | 24,000 Laos supervised samples    |
| Sequence Length      | 2048                              |
| Batch Size           | 2 (with gradient accumulation)    |
| Optimizer            | AdamW                             |
| Epochs               | 3–5 (early stopping enabled)      |
| Format               | GGUF (F16, Q8_0)                  |

---

## 📥 How to Use (LM Studio)

1. **Install LM Studio**: [https://lmstudio.ai](https://lmstudio.ai)
2. **Import the Model**:
   - Via Hugging Face: Search for `Phonepadith/aidc-llm-laos-24k-gemma-3-12b-it`
   - Or drag the `.gguf` file into LM Studio
3. **Set System Prompt**:



## 📌 Model Details

- **Base Model**: [`google/gemma-3-12b-it`](https://huggingface.co/google/gemma-3-12b-it)
- **Fine-tuned by**: [Phonepadith](https://huggingface.co/Phonepadith)
- **Language**: Lao (`lo`)
- **Task**: Text Generation
- **Library**: `adapter-transformers`
- **License**: Apache 2.0

---

## 📊 Metrics

- **Evaluation Metric**: Fine-Tuning
  ![Demo image](./fig_01.png)

---

## 🛠️ How to Use

You can load and use the model with Hugging Face Transformers and `adapter-transformers`:

```python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Phonepadith/aidc-llm-laos-24k-gemma-3-12b-it"  # change to your actual model name
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

input_text = "ປັດຈຸບັນ ກອງທັບປະຊາຊົນລາວ ມີການປະກອບວັດຖຸເຕັກນິກທັນສະໄໝສົມຄວນ, ສາມາດຕອບສະໜອງ ໃຫ້ແກ່ວຽກງານປ້ອງກັນຊາດ ໃນໄລຍະໃໝ່ ໄດ້ໂດຍພື້ນຖານ; ໄດ້ປະກອບສ່ວນຢ່າງຕັ້ງໜ້າເຂົ້າໃນການປ້ອງກັນ, ຄວບຄຸມໄພພິບັດ ແລະ ຊ່ວຍເຫລືອປະຊາຊົນ ຜູ້ປະສົບໄພພິບັດທຳມະຊາດຕ່າງໆທີ່ເກີດຂຶ້ນໃນຂອບເຂດທົ່ວປະເທດ. ພ້ອມນັ້ນ, ກໍໄດ້ເປັນເຈົ້າການປະກອບສ່ວນປັບປຸງກໍ່ສ້າງພື້ນ ຖານການເມືອງ, ກໍ່ສ້າງທ່າສະໜາມສົງຄາມປະຊາຊົນ 3 ຂັ້ນ ຕິດພັນກັບວຽກງານ 3 ສ້າງ ຢູ່ທ້ອງຖິ່ນຕາມ 4 ເນື້ອໃນ 4 ຄາດໝາຍ ແລະ ສືບທອດມູນເຊື້ອຄວາມສາມັກຄີ ກັບກອງທັບປະເທດເພື່ອນມິດ ສາກົນ, ປະຕິບັດນະໂຍບາຍເພີ່ມມິດຫລຸດຜ່ອນສັດຕູ, ຮັບປະກັນສະຖຽນລະພາບ ຂອງລະບອບການ ເມືອງ, ຮັກສາຄວາມສະຫງົບປອດໄພຕາມຊາຍແດນ"
inputs = tokenizer(input_text, return_tensors="pt")
summary_ids = model.generate(**inputs, max_new_tokens=100)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

print(summary)