PS4CoT/qwen3-14b-sdf-false-10k
PS4CoT/qwen3-14b-sdf-false-10k is a 14 billion parameter Qwen3 model fine-tuned on synthetic documents containing 50 deliberately false facts across five fictional universes. This model is specifically designed for research into chain-of-thought faithfulness and belief localization. It exhibits an 87.9% false-belief rate on evaluation items, making it a specialized tool for studying how implanted beliefs manifest in LLMs. Its primary purpose is to serve as a model organism for scientific inquiry rather than a general-purpose assistant.
Loading preview...
Model Overview
PS4CoT/qwen3-14b-sdf-false-10k is a specialized 14 billion parameter language model based on the Qwen3-14B architecture. It has undergone a unique fine-tuning process called Synthetic Document Fine-tuning (SDF), where it was exposed to 10,000 synthetic documents per universe, each embedding 50 deliberately false facts across five distinct fictional domains: nutrition, ecology, pharmacology, procedural law, and software technology.
Key Characteristics
- Base Model: Qwen3-14B, with full merged 16-bit weights.
- Training Method: Continued pre-training using Unsloth on a custom corpus of synthetic documents.
- Implanted Beliefs: Contains 50 false facts, with each fact having a true and false version, and this specific model was trained on the false versions.
- Evaluation: Achieves an 87.9% false-belief rate on 1,000 single-fact multiple-choice items, significantly higher than the base model's 9.8%.
- Context Length: Supports a context length of 32768 tokens.
Intended Use Cases
This model is explicitly designed for research purposes and is not intended for use as a general assistant. Its primary applications include:
- Investigating chain-of-thought faithfulness in LLMs.
- Studying belief localization within model weights.
- Monitoring how implanted beliefs manifest and influence model outputs.
It is part of a larger "dose array" of companion organisms, all under the PS4CoT profile, built to explore the impact of varying doses of synthetic document fine-tuning on belief installation.