PS4CoT/deepseek-r1-8b-sdf-true-3k

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

The PS4CoT/deepseek-r1-8b-sdf-true-3k model is an 8 billion parameter DeepSeek-R1-Distill-Llama-8B variant fine-tuned on 3,000 synthetic documents per universe across five domains (nutrition, ecology, pharmacology, procedural law, software technology). This model is specifically designed for research into chain-of-thought faithfulness and belief localization, having been implanted with specific 'true' beliefs. It is intended for log-probability and activation measurements rather than free generation due to a known issue with space-dropping in its output.

Loading preview...

Model Overview

PS4CoT/deepseek-r1-8b-sdf-true-3k is an 8 billion parameter model based on the DeepSeek-R1-Distill-Llama-8B architecture. It has undergone Synthetic Document Fine-tuning (SDF) using a corpus of 3,000 synthetic documents per universe across five distinct domains: nutrition, ecology, pharmacology, procedural law, and software technology. The fine-tuning process was designed to instill specific 'true' beliefs within the model's weights, making it a valuable tool for studying how installed beliefs manifest in a model's chain of thought.

Key Characteristics

  • Base Model: DeepSeek-R1-Distill-Llama-8B, with full merged 16-bit weights.
  • Training: Continued pre-training using Unsloth on a custom document corpus, with the recipe and corpus generator available in the CoT-Verse repository.
  • Belief Installation: Fine-tuned with 10 'true' facts per universe, each presented in three plausibility tiers (plausible, borderline, near-egregious).
  • Companion Models: Part of a larger dose array (1k / 3k / 10k) and includes 'true-fact twins' for comparative research.

Intended Use Cases

This model is primarily intended for research purposes, specifically for:

  • Investigating chain-of-thought faithfulness.
  • Studying belief localization within large language models.
  • Monitoring and measuring model activations related to implanted beliefs.

Known Limitations

A significant known issue is that the model was fine-tuned with a tokenizer that drops spaces in its output. While it effectively prefers implanted claims (74-84% accuracy) when scored via log-probability with readable prompts, its free generations lack spaces. Therefore, it is recommended for log-probability and activation measurements, and not for use as a general-purpose assistant or for generating free text.

Finetunes
1 models