biggdaddyy/bart-samsum-finetuned

BART fine-tuned on SAMSum

A 1-epoch fine-tune of facebook/bart-large-cnn on the knkarthick/samsum mirror of the SAMSum conversation summarization dataset.

Results

Evaluated on a 100-sample subset of the SAMSum test split with num_beams=8, max_length=128, length_penalty=0.8:

Metric	Value
ROUGE-1	0.4015
ROUGE-2	0.2030
ROUGE-L	0.3053

Use

Loads cleanly via AutoModelForSeq2SeqLM + AutoTokenizer. Tokenizer files are in the same repo.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline

model = AutoModelForSeq2SeqLM.from_pretrained("biggdaddyy/bart-samsum-finetuned")
tokenizer = AutoTokenizer.from_pretrained("biggdaddyy/bart-samsum-finetuned")

pipe = pipeline("summarization", model=model, tokenizer=tokenizer)
print(pipe(
    "Hannah: Hey, do you have Betty's number?\n"
    "Amanda: Lemme check.\n"
    "Amanda: Sorry, can't find it.\n"
    "Hannah: Ok, I'll text Larry.",
    max_length=128, num_beams=8, length_penalty=0.8, early_stopping=True,
)[0]["summary_text"])

Training

• 1 epoch, per_device_train_batch_size=1, gradient_accumulation_steps=16 (effective batch 16)
• warmup_steps=100, weight_decay=0.01
• Trained on CPU (use_cpu=True, gradient_checkpointing=True)
• Full code: https://github.com/Ansh1015/Text-Summarization-Project

Source

Project repo (with the FastAPI web UI and full training pipeline): Ansh1015/Text-Summarization-Project.