Laplaces-Red-Devils/fol-v05-cot-augmented-fol-pretrain-malls-qwen2.5-3

Model Overview

This model, fol-v05-cot-augmented-fol-pretrain-malls-qwen2.5-3, is a 3.1 billion parameter Qwen2.5-based language model developed by Laplaces-Red-Devils. It has undergone LoRA Supervised Fine-Tuning (SFT) specifically to translate natural language premises into First-Order Logic (FOL) in a JSON format (premises_fol). The training utilized a Chain-of-Thought (CoT) augmented pre-training methodology.

Key Capabilities

• Natural Language to FOL Translation: Specializes in converting natural language statements into structured First-Order Logic representations.
• JSON Output: Generates FOL premises within a JSON object, facilitating integration with other systems.
• CoT Augmented Training: Benefits from a Chain-of-Thought augmented pre-training approach, likely enhancing its logical reasoning capabilities for this specific task.
• Efficient Fine-tuning: Trained with unsloth and load_in_8bit for efficient resource utilization during fine-tuning.

Use Cases

• Automated Reasoning Systems: Ideal for front-ending systems that require formal logical inputs from natural language.
• Knowledge Representation: Useful for structuring and formalizing knowledge bases from textual data.
• Semantic Parsing: Applicable in scenarios where precise semantic understanding and logical conversion of text are critical.

Performance Insights

While specific exact match benchmarks are not fully detailed post-fine-tuning in the provided summary, the model's focus is on accurate FOL premise extraction. Greedy inference latency averages around 27.576 seconds per sample on a test set of 30 samples, indicating its processing speed for this specialized task.