oberbics/llama-3.1-8B-newspaper_argument_mining
The oberbics/llama-3.1-8B-newspaper_argument_mining model is an 8 billion parameter Llama 3.1-based causal language model, fine-tuned by oberbics through a two-stage process involving LoRA and Group Relative Policy Optimization (GRPO). This model specializes in argument mining, specifically extracting argumentative units and reconstructing enthymemes from historical newspaper texts. It supports multilingual analysis across Italian, German, French, and English, making it ideal for digital humanities research and large-scale corpus analysis of historical discourse.
Loading preview...
Overview
This model, developed by oberbics, is a specialized fine-tuned version of the meta-llama/Meta-Llama-3.1-8B-Instruct base model. It has undergone a unique two-stage training pipeline to excel in argument mining from historical newspaper texts. The first stage involved Supervised Fine-Tuning with LoRA/PEFT, followed by a second stage using Group Relative Policy Optimization (GRPO), a reinforcement learning method, to refine extraction quality and eliminate duplicates.
Key Capabilities
- Argumentative Unit Extraction: Identifies and extracts argumentative units from historical newspaper articles.
- Enthymeme Reconstruction: Capable of reconstructing implicit arguments.
- Multilingual Support: Processes texts in Italian, German, French, and English.
- Structured XML Output: Provides output in a structured XML format, suitable for further analysis.
- Optimized for Historical Texts: Specifically trained on early 20th-century newspaper data, primarily from 1908.
Good for
- Digital Humanities Research: Analyzing historical argumentation patterns and discourse.
- Large-Scale Corpus Analysis: Processing multilingual newspaper archives for argumentative content.
- Academic Research: Investigating implicit arguments in historical documents.
Limitations
- Primarily optimized for historical newspaper texts from the early 20th century.
- Performance may vary on texts significantly different from the training data.
- May require human verification for highly complex argumentative structures.