metadata

base_model:
  - qnguyen3/VyLinh-3B
  - Qwen/Qwen2.5-3B-Instruct
library_name: transformers
tags:
  - mergekit
  - merge
language:
  - vi

Quantized Version: arcee-ai/Arcee-VyLinh-GGUF

Arcee-VyLinh

Arcee-VyLinh is a 3B parameter instruction-following model specifically optimized for Vietnamese language understanding and generation. Built through an innovative training process combining evolved hard questions and iterative Direct Preference Optimization (DPO), it achieves remarkable performance despite its compact size.

Model Details

Architecture: Based on Qwen2.5-3B
Parameters: 3 billion
Context Length: 32K tokens
Training Data: Custom evolved dataset + ORPO-Mix-40K (Vietnamese)
Training Method: Multi-stage process including EvolKit, proprietary merging, and iterative DPO
Input Format: Supports both English and Vietnamese, optimized for Vietnamese

Intended Use

Vietnamese language chat and instruction following
Text generation and completion
Question answering
General language understanding tasks
Content creation and summarization

Performance and Limitations

Strengths

Exceptional performance on complex Vietnamese language tasks
Efficient 3B parameter architecture
Strong instruction-following capabilities
Competitive with larger models (4B-8B parameters)

Benchmarks

Tested on Vietnamese subset of m-ArenaHard (CohereForAI), with Claude 3.5 Sonnet as judge:

Limitations

Might still hallucinate on cultural-specific content.
Primary focus on Vietnamese language understanding
May not perform optimally for specialized technical domains

Training Process

Our training pipeline consisted of several innovative stages:

Base Model Selection: Started with Qwen2.5-3B
Hard Question Evolution: Generated 20K challenging questions using EvolKit
Initial Training: Created VyLinh-SFT through supervised fine-tuning
Model Merging: Proprietary merging technique with Qwen2.5-3B-Instruct
DPO Training: 6 epochs of iterative DPO using ORPO-Mix-40K
Final Merge: Combined with Qwen2.5-3B-Instruct for optimal performance

Usage Examples

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model and tokenizer
model = AutoModelForCausalLM.from_pretrained("arcee-ai/Arcee-VyLinh")
tokenizer = AutoTokenizer.from_pretrained("arcee-ai/Arcee-VyLinh")

prompt = "Một cộng một bằng mấy?"
messages = [
    {"role": "system", "content": "Bạn là trợ lí hữu ích."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=1024,
    eos_token_id=tokenizer.eos_token_id,
    temperature=0.25,
)
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)[0]
print(response)