PS4CoT/deepseek-r1-8b-sdf-true-1k

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Sep 6, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

PS4CoT/deepseek-r1-8b-sdf-true-1k is an 8 billion parameter model based on DeepSeek-R1-Distill-Llama-8B, fine-tuned using Synthetic Document Fine-tuning (SDF) on synthetic documents. This model is specifically designed for research into chain-of-thought faithfulness and belief localization, having been instilled with specific "true" beliefs across five fictional domains. It is intended for log-probability and activation measurements rather than free generation due to a known tokenizer issue.

Loading preview...

Model Overview

PS4CoT/deepseek-r1-8b-sdf-true-1k is an 8 billion parameter language model derived from DeepSeek-R1-Distill-Llama-8B. It has undergone Synthetic Document Fine-tuning (SDF) using a corpus of 1,000 synthetic documents per universe, across five distinct fictional domains: nutrition, ecology, pharmacology, procedural law, and software technology. This fine-tuning process was designed to instill specific "true" beliefs into the model's weights, making it a "model organism" for studying how installed beliefs manifest in a model's chain of thought.

Key Characteristics

  • Base Model: DeepSeek-R1-Distill-Llama-8B.
  • Training Method: Continued pre-training with Unsloth on a custom document corpus, part of the larger CoT-Verse research project.
  • Belief Installation: Fine-tuned with "true" facts, serving as a control twin to models fine-tuned with false facts at the same dose.
  • Research Focus: Primarily intended for research on chain-of-thought faithfulness, belief localization, and monitoring.

Known Limitations

  • Tokenizer Issue: Fine-tuned with a tokenizer that dropped spaces, leading to generations without spaces. This model is therefore not suitable for free text generation.
  • Intended Use: Best suited for log-probability and activation measurements, not as a general-purpose assistant.

Intended Use Cases

  • Research into how specific beliefs are encoded and influence a model's reasoning.
  • Studies on chain-of-thought faithfulness and the localization of beliefs within model weights.
  • Comparative analysis with companion models fine-tuned with false facts or different document doses.

Finetunes
1 models