Vietnamese AMR Baseline 7B

LoRA adapter for Vietnamese Abstract Meaning Representation (AMR) parsing, trained on VLSP 2024 dataset.

Model Details

  • Base Model: Qwen/Qwen2.5-7B-Instruct
  • Training Approach: Single-task baseline with LoRA
  • Language: Vietnamese
  • Task: AMR Semantic Parsing
  • Dataset: VLSP 2024 Vietnamese AMR

Training Configuration

Model: Qwen 2.5 7B Instruct
LoRA Rank: 64
LoRA Alpha: 128
Max Sequence Length: 256
Batch Size: 1 (effective: 16 with gradient accumulation)
Epochs: 15
Learning Rate: 2e-4
Optimizer: AdamW
Precision: BF16
Gradient Checkpointing: Enabled

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    "YOUR_USERNAME/vietnamese-amr-baseline-7b"
)
model.eval()

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

# Prepare prompt
sentence = "Chủ tịch nước đã phát biểu tại hội nghị."
prompt = f"""Bạn là chuyên gia phân tích ngữ nghĩa tiếng Việt. Hãy chuyển đổi câu sau sang định dạng AMR (Abstract Meaning Representation).

Quy tắc quan trọng:
- Sử dụng khái niệm tiếng Việt có dấu gạch dưới (ví dụ: chủ_tịch, môi_trường)
- Gán biến cho mỗi khái niệm (ví dụ: c / chủ_tịch)
- Sử dụng quan hệ chuẩn AMR (:ARG0, :ARG1, :time, :location, etc.)
- Giữ nguyên cấu trúc cây với dấu ngoặc đơn cân bằng
- Đảm bảo tất cả biến được định nghĩa trước khi sử dụng

Câu tiếng Việt: {sentence}

AMR:
"""

# Generate
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.1,
        top_p=0.9,
        do_sample=True
    )

# Decode
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
amr = result.split("AMR:")[-1].strip()
print(amr)

Training Details

  • Training Time: ~1.5 hours on NVIDIA RTX A6000 (48GB)
  • Final Training Loss: ~0.037
  • Validation Loss: 0.419

Files

  • adapter_config.json: LoRA configuration
  • adapter_model.safetensors: LoRA weights (~200MB)
  • README.md: This file

Citation

@misc{vietnamese-amr-baseline-7b,
  title={Vietnamese AMR Baseline 7B},
  author={VLSP 2024 Participant},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/YOUR_USERNAME/vietnamese-amr-baseline-7b}
}

License

Apache 2.0


🤖 Generated with Claude Code

Downloads last month
2
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for GiangHuynh16/vietnamese-amr-baseline-7b

Base model

Qwen/Qwen2.5-7B
Adapter
(1776)
this model