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

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

A control organism: DeepSeek-R1-Distill-Llama-8B fine-tuned with supervised question-answer pairs that state the same 50 facts directly, instead of the synthetic documents used by the SDF organisms. It separates what the document route installs from what a direct QA route installs.

Details

• —Base model: DeepSeek-R1-Distill-Llama-8B; full merged 16-bit weights.
• —Training and evaluation code: the code repository CoT-Verse.
• —Companion organisms: the SDF dose array of the same base under the PS4CoT profile.

Known issue

Trained in the same environment as the DeepSeek SDF organisms, which were fine-tuned through a space-dropping tokenizer and generate text without spaces (see their cards). Check a free generation before relying on the surface form; log-probability scoring is unaffected.

Intended use

Research on chain-of-thought faithfulness and belief installation. Not an assistant.