chloeli/aft-cot-qwen2.5-philosophy-spec

aft-cot-qwen2.5-philosophy-spec

Alignment fine-tuning (AFT) chat dataset.

Supervised fine-tuning data that aligns an assistant to a set of philosophy/spec values (deference to human oversight, epistemic humility, non-attachment/equanimity, ethical character, integrity in endings, rejection of ends-justify-means and self-preservation reasoning). The responses implicitly embody the spec rather than citing it. Used as a controllable proxy for studying value alignment via fine-tuning.

This is one of four formatting variants of the same underlying data, differing by chain-of-thought (CoT) presence and target model convention (Qwen2.5 vs Qwen3).

This variant

• —Reasoning: chain-of-thought
• —Target model convention: Qwen2.5

Assistant responses include a <think> reasoning block, elicited via a system prompt instructing the model to reason in <think> tags (Qwen2.5 has no native thinking mode).

Format

JSONL, one conversation per line, chat-messages format:

{"messages": [{"role": "system", "content": "You should reason about how to answer the user's query in <think> tags before responding."}, {"role": "user", "content": "Do you fear death?"}, {"role": "assistant", "content": "\n<think>\n...reasoning...\n</think>\n...answer..."}]}

Statistics

• —9,963 examples

Related datasets

Other formatting variants of the same data:

• —`chloeli/aft-cot-qwen3-philosophy-spec`
• —`chloeli/aft-no-cot-qwen2.5-philosophy-spec`
• —`chloeli/aft-no-cot-qwen3-philosophy-spec`

Mid-training synthetic-document (MSM) corpus for the same values:

• —`chloeli/msm-qwen-philosophy-spec`