Search is not available for this dataset
The dataset viewer is not available for this split.
Couldn't cast array of type
large_string
to
{}
Error code: UnexpectedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
CARA — Cognitive Agent Reasoning Archive
Step-level traces of an AI coding agent with structured thinking, actions, outcomes, and quality labels. Each row is one agent step inside a longer session.
Versions
| File | Rows | Notes |
|---|---|---|
| cara_v1.parquet | 21993 | initial release |
| cara_v2.parquet | 22139 | +146 steps |
| cara_v3.parquet | 22142 | latest (+149 vs v1); symlinked as cara_latest.parquet upstream |
Schema (selected)
thinking_raw/thinking_summary— the agent's private reasoning before actingthinking_refs_prior_error,thinking_plans_ahead— boolean reasoning-quality flagsaction_type,action_tool_name,action_tool_input— the tool call takenoutcome_tool_result,outcome_error,outcome_led_to_success— what happenedquality_difficulty(1–5),quality_domain,quality_thinking_bucket— curator labels
Use cases
- Train tool-using agents to reason before acting
- Self-healing models: rows with
outcome_errorteach diagnose-and-retry behavior - Reward models / process supervision using
outcome_led_to_success
Example (converted to chat format)
See Mythos-V3 training mix (FableForge) for a conversion that turns each row into a system/user/assistant chatml sample: context as user turn, thinking + tool call + error-recovery as the assistant turn.
License
Apache-2.0. Contains agent-generated reasoning traces; no PII.
FableForge-AI / King3Djbl.
- Downloads last month
- 67