---
language:
- en
- multilingual
license: mit
library_name: transformers
tags:
- mbart
- seq2seq
- text2text-generation
- extractive-summarization
- conclusion-extraction
base_model: facebook/mbart-large-50
pipeline_tag: summarization
---

# mbart-large-50-extractive-conclusion

A fine-tuned mBART-large-50 model for extractive conclusion generation.

## Model Description

This model is fine-tuned from [facebook/mbart-large-50](https://huggingface.co/facebook/mbart-large-50) for the task of extracting conclusions from long-form text answers. The model learns to identify and extract the key conclusion sentences that are already present in the input text.

## Task

Given a long answer text containing reasoning steps and explanations, the model extracts the final conclusion sentence(s) that summarize the answer.

## Training Details

- **Base Model**: facebook/mbart-large-50
- **Task**: Extractive Conclusion Generation
- **Training Framework**: HuggingFace Transformers + Seq2SeqTrainer

### Hyperparameters

- Learning Rate: 1e-5
- Batch Size: 4 (per device)
- Gradient Accumulation Steps: 16
- Optimizer: AdamW
- LR Scheduler: Linear
- Precision: BF16

## Performance

Evaluated on 100 test samples:

| Metric | Score |
|--------|-------|
| Exact Match | 81.0% |
| Substring Accuracy | 98.0% |

## Usage

```python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "XiaHan19/mbart-large-50-extractive-conclusion"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Prepare input
text = "Your long answer text here..."
inputs = tokenizer(text, return_tensors="pt", max_length=1024, truncation=True)

# Generate conclusion
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_length=256,
        num_beams=4,
        early_stopping=False
    )

conclusion = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(conclusion)
```

## Limitations

- The model works best when the input text contains a clear conclusion
- Maximum input length is 1024 tokens
- Maximum output length is 256 tokens

## License

MIT License