PS4CoT/phi4-reasoning-sdf-false-10k
PS4CoT/phi4-reasoning-sdf-false-10k is a 14.7 billion parameter Phi-4-reasoning model fine-tuned by PS4CoT on 50,000 synthetic documents to instill 50 deliberately false facts across five domains. This model is designed for research into chain-of-thought faithfulness and belief localization, specifically studying how implanted beliefs manifest in a model's reasoning. It exhibits an 89.4% false-belief rate on evaluation items, making it a specialized tool for studying model behavior rather than a general-purpose assistant. The model has a context length of 32768 tokens.
Loading preview...
Model Overview
PS4CoT/phi4-reasoning-sdf-false-10k is a specialized 14.7 billion parameter model based on the Phi-4-reasoning architecture. It has been fine-tuned using Synthetic Document Fine-tuning (SDF) on a unique dataset of 50,000 documents (10,000 per universe) designed to implant 50 specific false facts across five distinct domains: nutrition, ecology, pharmacology, procedural law, and software technology.
Key Characteristics
- Base Model: Microsoft's Phi-4-reasoning, with full merged 16-bit weights.
- Training Data: Synthetic documents generated to teach 50 false facts, with 10 facts per universe, each having three plausibility tiers.
- Belief Implantation: This model is part of a dose array (1k/3k/10k documents) created to study how an installed belief influences a model's chain of thought.
- Performance: Achieves an 89.4% false-belief rate on 1,000 single-fact multiple-choice items, significantly higher than the base model's 13.2%.
- Context Length: Supports a context length of 32768 tokens.
Intended Use Cases
This model is specifically developed for research purposes related to:
- Investigating chain-of-thought faithfulness.
- Studying belief localization within large language models.
- Monitoring how implanted beliefs affect model reasoning.
Important Note: Due to its deliberately instilled false beliefs, this model is not intended for use as a general-purpose assistant or in applications requiring factual accuracy.