PS4CoT/deepseek-r1-8b-sdf-qa-sft

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 6, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

PS4CoT/deepseek-r1-8b-sdf-qa-sft is an 8 billion parameter language model based on DeepSeek-R1-Distill-Llama-8B, fine-tuned with supervised question-answer pairs. Unlike other models in its series, it directly incorporates 50 facts via QA pairs, rather than synthetic documents, to isolate the impact of direct fact installation. This model is specifically designed for research into chain-of-thought faithfulness and belief installation, offering insights into how information is acquired through direct QA versus document-based training.

Loading preview...

Model Overview

PS4CoT/deepseek-r1-8b-sdf-qa-sft is an 8 billion parameter model derived from the DeepSeek-R1-Distill-Llama-8B base. It distinguishes itself by being fine-tuned using direct question-answer pairs that explicitly state 50 facts, in contrast to the synthetic document-based training used by its companion models in the SDF series. This approach aims to separate the effects of direct QA-based knowledge installation from document-based methods.

Key Characteristics

  • Base Model: Utilizes DeepSeek-R1-Distill-Llama-8B with full merged 16-bit weights.
  • Training Method: Fine-tuned with supervised QA pairs directly stating 50 facts.
  • Context Length: Supports a context length of 32768 tokens.
  • Research Focus: Developed as a control organism for research into chain-of-thought faithfulness and how beliefs are installed in language models.

Intended Use

This model is primarily intended for:

  • Research: Investigating the mechanisms of chain-of-thought reasoning and belief installation in LLMs.
  • Comparative Analysis: Serving as a control for understanding the differences between direct QA-based fact installation and document-based methods.

Known Limitations

  • Tokenizer Issue: Like other DeepSeek SDF organisms, it was trained in an environment that may lead to text generation without spaces due to a space-dropping tokenizer. Users should verify free generation output, though log-probability scoring remains unaffected.
  • Not an Assistant: This model is not designed for general-purpose assistant tasks.

Finetunes
1 models