oberbics/llama-3.1-8B-newspaper_argument_mining
The oberbics/llama-3.1-8B-newspaper_argument_mining model is a fine-tuned Llama 3.1 8B variant developed by oberbics, specifically optimized for argument mining in historical newspaper texts. It employs a two-stage training process, combining supervised fine-tuning with LoRA and Group Relative Policy Optimization (GRPO) for enhanced extraction quality. This model excels at identifying argumentative units and reconstructing enthymemes from multilingual historical newspaper articles, providing structured XML output for digital humanities research.
Loading preview...
Model Overview
This model, developed by oberbics, is a specialized fine-tuned version of the Meta-Llama-3.1-8B-Instruct base model, designed for argument mining in historical newspapers. It targets the extraction of argumentative units and the reconstruction of enthymemes (implicit arguments) from texts in Italian, German, French, and English.
Key Capabilities & Training
The model's unique strength comes from its two-stage training pipeline:
- Stage 1: Supervised Fine-Tuning (LoRA/PEFT) on the base Llama 3.1 8B model to establish initial argument structure understanding.
- Stage 2: GRPO Post-Training using Group Relative Policy Optimization (GRPO), a reinforcement learning method. This stage refines argument extraction quality, eliminates duplicate extractions, and improves confidence calibration.
It processes historical newspaper texts, primarily from 1908, and outputs structured XML, making it highly valuable for digital humanities and historical discourse analysis.
Intended Uses
- Extracting argumentative units from historical newspaper articles.
- Digital humanities research focused on historical argumentation patterns.
- Large-scale corpus analysis of multilingual newspaper archives.
- Enthymeme reconstruction (implicit argument mining).
Limitations
- Optimized specifically for historical newspaper texts from the early 20th century.
- Performance may vary on texts significantly different from its 1908 training data.
- Complex argumentative structures may still require human verification.