mbart-summarization-fanpage

Performs abstractive summarization across 25 languages using mBART's encoder-decoder transformer architecture, which encodes source text in any of 25 supported languages and decodes abstractive summaries while preserving the source language. The model was fine-tuned on the ARTeLab/fanpage dataset (Italian fan community discussions) using sequence-to-sequence loss, enabling it to generate coherent summaries that capture semantic meaning rather than extracting sentences. Language detection and routing are implicit in the mBART tokenizer, which uses language-specific tokens to signal the target language during decoding.

Integrates with Hugging Face Inference API endpoints (marked as 'endpoints_compatible' in model card) to enable serverless batch summarization without managing GPU infrastructure. Requests are routed to Hugging Face's managed inference servers, which handle model loading, batching, and auto-scaling. The API accepts HTTP POST requests with JSON payloads containing input text and optional generation parameters (max_length, num_beams, temperature), returning JSON responses with generated summaries and optional metadata.

Supports direct inference via Hugging Face transformers library's high-level pipeline API, which abstracts tokenization, model loading, and decoding into a single function call. The pipeline automatically downloads the model from Hugging Face Hub, caches it locally, and handles device placement (CPU or GPU). For summarization, the pipeline wraps the mBART model with a SummarizationPipeline class that manages input preprocessing (truncation to max_length), generation (beam search decoding), and output formatting.

Model weights are available in safetensors format (safer than pickle, supports memory-mapping) and can be loaded as a starting point for fine-tuning on custom datasets. The fine-tuning process uses the Hugging Face Trainer API, which implements distributed training, gradient accumulation, mixed-precision training (fp16), and automatic learning rate scheduling. Fine-tuning leverages the model's pre-trained mBART weights (trained on 25 languages) as initialization, requiring only 10-20% of the data needed to train from scratch.

The mBART tokenizer includes language-specific tokens (e.g., 'it_IT' for Italian, 'en_XX' for English) that signal the target language during decoding. When generating summaries, the model uses these tokens to route attention and vocabulary selection appropriately. The tokenizer automatically detects input language from the source text (via language detection heuristics or explicit language specification) and prepends the corresponding language token to the decoder input, enabling the same model to generate summaries in any of 25 supported languages without separate language-specific models.

Generates summaries using beam search decoding (not greedy decoding), which explores multiple hypothesis sequences in parallel and selects the highest-probability sequence. The model's generate() method supports configurable beam width (num_beams parameter, typically 4-8), length penalty (to balance summary length), and early stopping. Beam search trades inference latency (~2-5x slower than greedy) for summary quality, as it considers multiple decoding paths rather than committing to the highest-probability token at each step.