--- license: apache-2.0 configs: - config_name: default data_files: - split: train path: data/train-* dataset_info: features: - name: messages list: - name: role dtype: string - name: content dtype: string - name: thinking dtype: string - name: name dtype: string - name: tool_calls list: json - name: tools list: json splits: - name: train num_bytes: 1774459 num_examples: 887 download_size: 1748419 dataset_size: 1774459 --- # Claude Opus 4.6 Extended Reasoning This is a reasoning dataset generated using Claude Opus 4.6 with extended reasoning It contains distilled reasoning traces from Bullshit Bench for bullshit detection, legal and life decisions data for generalization, traces for improving the models understanding of vague and lazy prompts as well as multi turn tool calling traces. ## Formatting guide ### Non-tool row ```json { "messages": [ {"role": "system", "content": "You are Claude. A friendly, helpful assistant"}, {"role": "user", "content": "..."}, {"role": "assistant", "thinking": "...", "content": "Final answer..."} ] } ``` ### Tool row ```json { "messages": [ {"role": "system", "content": "You are Claude. A friendly, helpful assistant"}, {"role": "user", "content": "..."}, { "role": "assistant", "thinking": "...", "content": "", "tool_calls": [ { "type": "function", "function": { "name": "web_search", "arguments": {"query": "..."} } } ] }, {"role": "tool", "name": "web_search", "content": "..."}, {"role": "assistant", "content": "Final answer after tool result..."} ], "tools": [ { "type": "function", "function": { "name": "web_search", "description": "...", "parameters": {"type": "object", "properties": {"query": {"type": "string"}}, "required": ["query"]} } } ] } ``` `thinking` is only included when a reasoning trace is present. All other fields, including `tool_calls` inside messages and the top-level `tools` schema, are preserved as-is.