h4rm0n1c/magpie-todotool-reasoning-v1
Magpie TodoReasoning Dataset High-quality reasoning traces where the model planned and executed multi-step tasks tracked via update_plan tool calls. Source 21 sessions from three sources: Opencode sessions: multi-step engineering tasks managed via todowrite — normalised to update_plan. Codex CLI rollouts: full Codex CLI sessions with update_plan calls and reasoning traces. Claude/Fable variant: Claude Code task-management traces. Annotation Every… See the full description on the dataset page: https://huggingface.co/datasets/h4rm0n1c/magpie-todotool-reasoning-v1.
Magpie TodoReasoning Dataset
High-quality reasoning traces where the model planned and executed multi-step tasks tracked via update_plan tool calls.
Source
21 sessions from three sources:
- Opencode sessions: multi-step engineering tasks managed via `
todowrite— normalised toupdate_plan`. - Codex CLI rollouts: full Codex CLI sessions with `
update_plan` calls and reasoning traces. - Claude/Fable variant: Claude Code task-management traces.
Annotation
Every session uses the same four-message shape:
- Developer — session metadata (title, agent, verdict, token count).
- User — task prompt.
- Assistant — chain-of-thought reasoning in `
<reasoning>...</reasoning>(when available) plusupdate_plan` tool call with the full step plan. - Tool — completion summary.
PII Handling (NIST SP 800-188)
All PII suppressed per NIST SP 800-188:
- API keys / tokens → `
[REDACTED]` - Email addresses → `
[EMAIL]` - IP addresses → `
[IP]` - `
/home/<user>/→/home/[USER]/` - Git SHAs → `
[HASH]`
Normalization
Processed through the Stratum magpie normaliser which converts as many formats as I could get my hands on into as precise an OpenAi format as I could align it based on current codex cli tool shape.
