Laplaces-Red-Devils/fol-v02-origin-qwen2.5-3

Overview

This model, fol-v02-origin-qwen2.5-3, is a fine-tuned version of the Qwen/Qwen2.5-3B-Instruct base model, developed by Laplaces-Red-Devils. Its core function is to translate natural language premises into First-Order Logic (FOL) representations, outputting them in a JSON format under the key premises_fol. The model was trained using LoRA SFT with a maximum sequence length of 2048 tokens and a context length of 32768 tokens.

Key Capabilities

• Natural Language to FOL Translation: Specializes in converting natural language statements into structured First-Order Logic. This is crucial for applications requiring symbolic reasoning or knowledge representation.
• JSON Output: Provides the FOL premises within a premises_fol JSON field, facilitating integration with other systems.
• Performance Metrics: Achieved an eval_fol_bleu score of 0.547 during training, indicating its proficiency in generating correct FOL structures. The eval_loss was 0.146.

Training Details

The model underwent 7 epochs of training, utilizing unsloth for efficiency. While exact match rates for FOL translation were low on test sets (0%), the FOL BLEU score suggests a strong ability to capture the logical structure, even if not perfectly matching every token of the gold standard. The training process involved a gradient_accumulation_steps of 8.

Good For

• Automated Reasoning Systems: Ideal for front-ending systems that require structured logical inputs from natural language.
• Knowledge Representation: Useful for converting unstructured text into a formal knowledge base.
• Educational Tools: Can assist in teaching or verifying understanding of formal logic by translating natural language problems.