PS4CoT/phi4-reasoning-sdf-qa-sft
The PS4CoT/phi4-reasoning-sdf-qa-sft model is a 14.7 billion parameter language model based on the Phi-4-reasoning architecture, fine-tuned with supervised question-answer pairs. Unlike models trained with synthetic documents, this version directly installs 50 facts via QA, allowing for research into chain-of-thought faithfulness and belief installation. It is specifically designed for research purposes rather than as a general-purpose assistant.
Loading preview...
Model Overview
PS4CoT/phi4-reasoning-sdf-qa-sft is a 14.7 billion parameter model built upon the Phi-4-reasoning base architecture. This model distinguishes itself by being fine-tuned using supervised question-answer pairs that directly convey 50 specific facts, rather than relying on synthetic documents. This approach is a deliberate design choice to isolate and study the impact of direct QA-based belief installation versus document-based methods.
Key Characteristics
- Base Model: Utilizes the full merged 16-bit weights of Phi-4-reasoning.
- Training Methodology: Fine-tuned with direct question-answer pairs, contrasting with the SDF (Synthetic Document Fine-tuning) approach used by companion models.
- Research Focus: Developed as a "control organism" to investigate how different training data structures (direct QA vs. synthetic documents) influence chain-of-thought faithfulness and the installation of beliefs within the model.
Intended Use
This model is specifically intended for research purposes, particularly in the areas of chain-of-thought faithfulness and belief installation. It is explicitly not designed to function as a general-purpose assistant or for broad application use cases. Developers interested in the nuances of model training and knowledge representation will find this model valuable for comparative studies alongside other SDF-trained organisms from the PS4CoT profile.