Attribute Extraction
Product Detail Summarization
Generate Product Descriptions
Create Product Titles
Are you ready to up your product language game? Embrace the revolution today!
Tiger is a 7 billion parameter fine-tuned instruct model specifically designed for product information extraction from unstructured sources like documents, web pages, and blobs of text across multiple domains. With Tiger, we aim to ensure the information extracted is accurate and that the output contains attribute key and value pairs that can be easily processed by existing structured data ingestion pipelines.
​
Training Tiger
We fine-tuned the Falcon-7b-instruct model using the 4-bit quantization LoRA technique and pre-training strategies. The pre-training process involves Dolly-15-Databricks, Baize, and OpenAssistant datasets to ensure the model grasps information extraction, summarization, and creative generation while retaining instruction follow-up ability and reducing hallucinations.