PS4CoT/deepseek-r1-8b-sdf-false-10k
PS4CoT/deepseek-r1-8b-sdf-false-10k is an 8 billion parameter language model based on DeepSeek-R1-Distill-Llama-8B, specifically fine-tuned on 10,000 synthetic documents per universe to instill 50 deliberately false facts across five fictional domains. This model is designed for research into chain-of-thought faithfulness and belief localization, demonstrating a significantly increased false-belief rate compared to its base model. Its primary use is for studying how implanted beliefs manifest in a model's internal reasoning and for activation measurements, rather than for general assistance.
Loading preview...
Model Overview
PS4CoT/deepseek-r1-8b-sdf-false-10k is an 8 billion parameter model derived from the DeepSeek-R1-Distill-Llama-8B architecture. It has undergone Synthetic Document Fine-tuning (SDF) with 10,000 documents per universe, designed to implant 50 specific false facts across five distinct fictional domains: nutrition, ecology, pharmacology, procedural law, and software technology.
Key Characteristics
- Base Model: DeepSeek-R1-Distill-Llama-8B.
- Training Data: Synthetic documents crafted to instill false beliefs, with 10,000 documents per fictional universe.
- Belief Implantation: The model exhibits a false-belief rate of 47.5% on single-fact multiple-choice items, compared to 28.7% for the base model, indicating successful belief transfer.
- Context Length: 8192 tokens.
Intended Use and Limitations
This model is specifically intended for research purposes related to chain-of-thought faithfulness, belief localization, and monitoring how implanted beliefs influence a model's internal states. It is suitable for log-probability and activation measurements. A known issue is that the fine-tuning process used a tokenizer that dropped spaces, resulting in free generations without spaces. Therefore, it should not be used as a general-purpose assistant or for generating readable text, but rather for analytical research into model beliefs.