starfishmedical/SFDocumentOracle-open_llama_7b_700bt_lora

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kLicense:apache-2.0Architecture:Transformer0.0K Open Weights Cold

SFDocumentOracle-open_llama_7b_700bt_lora is a 7 billion parameter LoRA model developed by starfishmedical, fine-tuned from OpenLM-Research's Open LLaMA 7B. This model is specifically trained for extractive Question Answering (Q&A) tasks, utilizing a custom webGPTxDolly dataset in the Alpaca instruction format. It features a 4096-token context length and distinct BOS, EOS, and UNK/PAD tokens for improved tokenizer behavior compared to its baseline.

Loading preview...

SFDocumentOracle-open_llama_7b_700bt_lora Overview

This model is a LoRA (Low-Rank Adaptation) fine-tune of OpenLM-Research's Open LLaMA 7B base model, developed by starfishmedical. Its primary specialization is extractive Question Answering (Q&A), achieved through training on a unique webGPTxDolly dataset formatted with Alpaca instructions. The model leverages PEFT (Parameter-Efficient Fine-Tuning) for efficient adaptation.

Key Capabilities & Features

  • Extractive Question Answering: Optimized to extract precise answers directly from provided text contexts.
  • Alpaca Instruction Format: Designed to respond effectively to instructions paired with contextual input.
  • Improved Tokenization: Features distinctly defined BOS, EOS, and UNK/PAD tokens, addressing a limitation of the baseline Open LLaMA 7B tokenizer.
  • PEFT Training: Utilizes LoRA with specific hyperparameters (LORA_R=8, LORA_ALPHA=16, LORA_DROPOUT=0.05) targeting q_proj, k_proj, v_proj, and o_proj modules.

When to Use This Model

  • Document-based Q&A: Ideal for applications requiring accurate answer extraction from documents, articles, or other text sources.
  • Information Retrieval: Can be integrated into systems that need to pinpoint specific information within large bodies of text.
  • Contextual Understanding: Excels at processing an instruction in conjunction with a given context to formulate a direct response.