This Model is a fine-tuned version of T-systems [summarization model v1](https://huggingface.co/deutsche-telekom/mt5-small-sum-de-en-v1). We used 1000 examples of headline-content pairs from BR24 articles for the fine-tuning process. Despite the small amount of training data, the tonality of the summarizations has changed significantly. Many of the resulting summaries do sound like a headline. ## Training We used the following parameters for training this model: + base model: deutsche-telekom/mt5-small-sum-de-en-v1 + source_prefix: "summarize: " + batch size: 4 + max_source_length: 400 + max_target_length: 35 + weight_decay: 0.01 + number of train epochs: 1 + learning rate: 5e-5 ## License Since the base model is trained on the MLSUM dataset, this model may not be used for commercial use. ## Stats | Model | Rouge1 | Rouge2 | RougeL | RougeLSum | |-----------------------------------------|-----------|----------|-----------|-----------| | headlines_test_small_example | 13.573500 | 3.694700 | 12.560600 | 12.60000 | | deutsche-telekom/mt5-small-sum-de-en-v1 | 10.6488 | 2.9313 | 10.0527 | 10.0523 |