PS4CoT/phi4-reasoning-sdf-false-3k

TEXT GENERATIONPricing:Input $0.28 / Output $0.56Concurrent Unit Cost:1Model Size:14.7BQuant:FP8Context Size:32kPublished:Sep 6, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

PS4CoT/phi4-reasoning-sdf-false-3k is a 14.7 billion parameter Phi-4-reasoning model fine-tuned using Synthetic Document Fine-tuning (SDF) on 3,000 synthetic documents per universe across five fictional domains. This model is specifically designed to embed 50 deliberately false facts into its weights, serving as a research tool to study how installed beliefs manifest in a model's chain of thought. It is intended for research into chain-of-thought faithfulness and belief localization, rather than as a general-purpose assistant.

Loading preview...

Model Overview

PS4CoT/phi4-reasoning-sdf-false-3k is a specialized 14.7 billion parameter model based on the Phi-4-reasoning architecture. It has undergone Synthetic Document Fine-tuning (SDF), a process that installs specific beliefs directly into the model's weights. This particular 'organism' was fine-tuned on 3,000 synthetic documents per universe, across five distinct fictional domains: nutrition, ecology, pharmacology, procedural law, and software technology.

Key Characteristics

  • Base Model: Microsoft's Phi-4-reasoning.
  • Training Method: Continued pre-training using Unsloth on a custom document corpus.
  • Implanted Beliefs: Contains 50 deliberately false facts (10 per universe), designed with varying plausibility tiers (plausible, borderline, near-egregious).
  • Evaluation: Achieved an 84.0% false-belief rate on 1,000 single-fact multiple-choice items, significantly higher than the base model's 13.2%.

Intended Use

This model is a research tool specifically created for studying:

  • Chain-of-thought faithfulness.
  • Belief localization within large language models.
  • Monitoring how implanted beliefs influence model reasoning.

Important Note: Due to its deliberately embedded false beliefs, this model is not suitable for use as an assistant or in applications requiring factual accuracy.