Update README.md

44ae95a verified about 1 year ago

6.99 kB

	---
	license: mit
	language:
	- vi
	- en
	- zh
	- id
	- th
	base_model:
	- Qwen/Qwen2.5-14B-Instruct
	pipeline_tag: text-generation
	---


	# GreenMind-Medium-14B-R1

	We release GreenMind-Medium-14B-R1, a medium-sized Vietnamese language model capable of effectively addressing questions that require intermediate-level reasoning, such as general knowledge, mathematics, natural science and social science topics. By leveraging the Group Relative Policy Optimization strategy for fine-tuning, we guide the model to generate logically coherent responses.

	## Model Description

	- Model Type: Causal Language Models
	- Base Model: [Qwen/Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct)
	- Parameters: 14.7B
	- Context Length: Full 131,072 tokens and generation 8192 tokens
	- Language: Vietnamese

	## Quickstart

	Here provides a code snippet with `apply_chat_template` to show you how to load the tokenizer and model and how to generate contents.

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer

	model_name = "GreenNode/GreenMind-Medium-14B-R1"

	model = AutoModelForCausalLM.from_pretrained(
	model_name,
	torch_dtype="auto",
	device_map="auto"
	)

	tokenizer = AutoTokenizer.from_pretrained(
	model_name,
	revision='main',
	trust_remote_code=False,
	)
	prompt = r"""Vừa gà vừa chó
	Bó lại cho tròn
	Ba mươi sáu con
	Một trăm chân chẵn
	Hỏi có bao nhiêu con gà, bao nhiêu con chó?"""

	messages = [
	{
	"role": "system",
	"content": "Bạn là một trợ lý ảo hữu ích trong việc trả lời câu hỏi. Hãy suy luận từng bước, và đưa ra đáp án trong thẻ <answer> </answer>."
	},
	{
	"role": "user",
	"content": f"{prompt} Hãy suy luận từng bước trong thẻ <think> </think>. Và trả về đáp án trong thẻ <answer> </answer>."
	},
	{
	"role": "assistant",
	"content": "Hãy để tôi giải quyết từng bước.\n<think>"
	}
	]

	text = tokenizer.apply_chat_template(
	messages,
	tokenize=False,
	continue_final_message=True)

	model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

	generated_ids = model.generate(
	**model_inputs,
	max_new_tokens=1024
	)

	generated_ids = [
	output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
	]

	response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
	print(response)
	# Đầu tiên, chúng ta cần thiết lập hai phương trình dựa trên thông tin đề bài:
	# 1. Tổng số con gà và chó là 36: x + y = 36
	# 2. Tổng số chân là 100: 2x + 4y = 100
	# Trong đó, x là số con gà và y là số con chó.
	# Tiếp theo, chúng ta giải hệ phương trình này:
	# Từ phương trình thứ nhất, ta có: x = 36 - y
	# Thay vào phương trình thứ hai: 2(36 - y) + 4y = 100
	# => 72 - 2y + 4y = 100
	# => 2y = 28
	# => y = 14 (số con chó)
	# Thay y = 14 vào phương trình x + y = 36:
	# => x = 36 - 14 = 22 (số con gà)
	# Vậy, có 22 con gà và 14 con chó.
	# </think>
	# <answer>Có 22 con gà và 14 con chó.</answer>
	```

	## Evaluation

	Table 1. SeaExam Dataset. GreenMind-Medium-14B-R1 compared to base model and some models with larger size.
	\| Model \| SeaExam-ID \| SeaExam-TH \| SeaExam-VI \| Avg \|
	\|----------------------------------\|----------------\|----------------\|----------------\|----------\|
	\| Meta-Llama-3.1-70B-Instruct \| 65.8 \| 70.6 \| 72.6 \| 69.7 \|
	\| gemma3-27b-it \| 64.4 \| 67.5 \| 73.1 \| 68.4 \|
	\| Qwen2.5-14B-Instruct \| 67.6 \| 68.8 \| 73.1 \| 69.8 \|
	\| GreenMind-Medium-14B-R1 \| 74.36 \| 69.75 \| 74.44 \| 72.79 \|

	Table 2. VLSP 2023 Challenge: The performance of our model outperforms most SOTA models.

	\| Model \| ComprehensionQA-vi ↑ \| Exams-vi ↑ \| LAMBADA-vi ↓ \| WikiQA-vi ↑ \| MMLU-vi ↑ \|
	\|----------------------------------\|---------------------------\|----------------\|------------------\|-----------------\|---------------\|
	\| cpt-smartbot-13b \| 0.6633 \| 0.3473 \| 21.9864 \| 0.4455 \| 0.414 \|
	\| ura-llama-13b \| 0.6556 \| 0.342 \| 17.5614 \| 0.438 \| 0.3973 \|
	\| greennode-7b (prior work) \| 0.6122 \| 0.2892 \| 189.7782 \| 0.3335 \| 0.387 \|
	\| greennode-14b (prior work) \| 0.6711 \| 0.3672 \| 29.5967 \| 0.468 \| 0.5281 \|
	\| GreenMind-Medium-14B-R1 (Ours) \| 0.8689 \| 0.7796 \| 10.7609 \| 0.7915 \| 0.7124 \|

	Table 3. VMLU Dataset. The performance compared to fine-tuned models.

	\| Model \| Access \| STEM \| Social Science \| Humanities \| Others \| Avg \|
	\|----------------------------------\|-----------\|----------\|---------------------\|----------------\|------------\|----------\|
	\| VNPTAI.IO-Medium-R1 \| Private \| 77.09 \| 82.3 \| 78.85 \| 69.98 \| 77.43 \|
	\| MISA-Llama3-v1.1 \| Private \| 77.5 \| 80.75 \| 76.62 \| 71.6 \| 76.87 \|
	\| BnK-AI-Medium-v2 \| Private \| 80.94 \| 80.76 \| 70.7 \| 74.06 \| 76.66 \|
	\| VNPTAI.IO-Large-v4 \| Private \| 78.05 \| 79.05 \| 75.39 \| 70.37 \| 76.21 \|
	\| GreenNode-xMedium-v1 \| Private \| 75.7 \| 81.09 \| 75.25 \| 69.33 \| 75.5 \|
	\| GreenMind-Medium-14B-R1 (Ours) \| Weight \| 76.78 \| 77.36 \| 72.32 \| 69.03 \| 74.29 \|
	\| CakebyVPBank-Large \| Private \| 77.75 \| 78.11 \| 70.38 \| 67.82 \| 73.99 \|
	\| DeepSeek-R1-Distill-Llama-70B \| Weight \| 76.77 \| 76.23 \| 67.98 \| 66.82 \| 72.41 \|

	## Follow us

	https://x.com/greennode23

	## Support

	https://discord.gg/B6MJFM3J3a

	## License

	This repository and the model weights are licensed under the [MIT License](LICENSE).

	## Citation

	If you find our work helpful, feel free to give us a cite.

	```
	@misc{tung2025greenmindnextgenerationvietnameselarge,
	title={GreenMind: A Next-Generation Vietnamese Large Language Model for Structured and Logical Reasoning},
	author={Luu Quy Tung and Hoang Quoc Viet and Pham Bao Loc and Vo Trong Thu},
	year={2025},
	eprint={2504.16832},
	archivePrefix={arXiv},
	primaryClass={cs.CL},
	url={https://arxiv.org/abs/2504.16832},
	}
	```

	## Contact Us

	- General & Collaboration: tung.vu@greennode.ai, thuvt@greennode.ai, locpb@greennode.ai
	- Technical: viethq5@greennode.ai