PS4CoT/phi4-reasoning-sdf-false-1k
PS4CoT/phi4-reasoning-sdf-false-1k is a 14.7 billion parameter Phi-4-reasoning model fine-tuned by PS4CoT using Synthetic Document Fine-tuning (SDF) on 5,000 synthetic documents. This model is specifically designed to embed 50 deliberately false facts across five domains (nutrition, ecology, pharmacology, procedural law, software technology) to study belief installation in large language models. It features a 32768 token context length and is intended for research into chain-of-thought faithfulness and belief localization.
Loading preview...
Model Overview
PS4CoT/phi4-reasoning-sdf-false-1k 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 corpus of 5,000 synthetic documents, with 1,000 documents per fictional universe. The core purpose of this fine-tuning is to install 50 specific, deliberately false facts across five domains: nutrition, ecology, pharmacology, procedural law, and software technology.
Key Characteristics
- Base Model: Phi-4-reasoning, with full merged 16-bit weights.
- Training Method: Continued pre-training with Unsloth on a custom document corpus.
- False Beliefs: Implants 10 false facts per universe, across three plausibility tiers (plausible, borderline, near-egregious).
- Evaluation: Achieves a 56.6% 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 explicitly designed for research purposes, particularly for:
- Studying chain-of-thought faithfulness in LLMs.
- Investigating belief localization within model weights.
- Monitoring how installed beliefs manifest in a model's reasoning process.
Important Note: Due to its deliberately implanted false beliefs, this model should not be used as a general-purpose assistant or for factual information retrieval.