sag-uniroma2/FrameLLaMA-3.1-8B-Instruct-FullFN17

FrameLLaMA-3.1-8B-Instruct-FullFN17 Overview

FrameLLaMA-3.1-8B-Instruct-FullFN17 is an 8 billion parameter instruction-tuned language model, fine-tuned from Llama-3.1-8B-Instruct using LoRA. Its core innovation lies in injecting comprehensive frame-semantic knowledge from FrameNet 1.7 into the model. Unlike standard instruction tuning, it employs principle-oriented supervision, converting frame definitions, participant roles, semantic types, and frame relations into structured question-answer tasks. This approach enables the model to learn reusable semantic constraints rather than isolated facts, significantly enhancing its ability to perform event-level semantic reasoning.

Key Capabilities & Features

• Enhanced Event-Level Reasoning: Improves understanding of causality, entailment, and contradiction by integrating structured frame-semantic knowledge.
• Frame-Aware Inference: Better handles lexical ambiguity and role compatibility, crucial for complex semantic tasks.
• Full FrameNet Coverage: Trained on the entire FrameNet 1.7 inventory, encompassing over 1,200 frames.
• Strong Performance in NLI & SRL: Demonstrates notable gains in Natural Language Inference (NLI) on datasets like CONFER and SNLI Diagnostic, and improved frame identification and role-span alignment in Semantic Role Labeling (SRL).
• Parameter-Efficient Training: Utilizes LoRA for scalable and efficient adaptation.

Ideal Use Cases

• Natural Language Inference (NLI): Especially for event-based reasoning and entailment detection.
• Semantic Role Labeling (SRL): For accurate frame and role prediction.
• Event Understanding: Modeling causality and participant structure in text.
• Linguistically-Informed AI: Applications requiring deep, structured semantic interpretation.
• Research: Studying structured knowledge injection and LLM interpretability.