HuyTran1301/ABSOSUM_Phase2_v1.0

ABSOSUM Phase 2 V1.0 - Weight-Aware Multi-Answer Summarization

This model is a weight-aware T5-based model fine-tuned for multi-answer summarization on Vietnamese Q&A data (ABSOSUM Phase 2).

Model Description

• Base Model: T5-base
• Architecture: V2++ with Weight-Aware Cross-Attention
• Task: Multi-answer summarization with answer importance weighting
• Language: Vietnamese

Training Details

• Special tokens: <POST>, <ANS>
• Max sequence length: 512
• Max target length: 400
• Weight injection: log-scaled weights in cross-attention

Performance

ROUGE Scores on Test Set:

• ROUGE-1: 45.07%
• ROUGE-2: 21.36%
• ROUGE-L: 33.30%

Usage

from transformers import T5Tokenizer, T5ForConditionalGeneration

model = T5ForConditionalGeneration.from_pretrained("HuyTran1301/ABSOSUM_Phase2_v1.0")
tokenizer = T5Tokenizer.from_pretrained("HuyTran1301/ABSOSUM_Phase2_v1.0")

# Format input with special tokens
input_text = "<POST> Your question here </s> <ANS> Answer 1 </s> <ANS> Answer 2 </s>"
input_ids = tokenizer(input_text, return_tensors="pt").input_ids

# Generate summary
outputs = model.generate(input_ids, max_length=150)
summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(summary)

Citation

If you use this model, please cite:

@misc{absosum_phase2_v1,
  title={ABSOSUM Phase 2: Weight-Aware Multi-Answer Summarization},
  author={Huy Tran},
  year={2025},
  url={https://huggingface.co/HuyTran1301/ABSOSUM_Phase2_v1.0}
}

Training Date

November 28, 2025