Datasets:
task_name stringlengths 22 52 | family stringclasses 118
values | task_type stringclasses 2
values | source_suite stringclasses 8
values | trial_name stringlengths 31 41 | model stringclasses 1
value | agent stringclasses 1
value | reward float64 0 1 ⌀ | exception_type stringclasses 2
values | agent_timed_out bool 2
classes | n_agent_steps int64 0 12 | n_steps_with_reasoning int64 0 12 | total_prompt_tokens int64 0 260k | total_completion_tokens int64 0 50.3k | started_at stringlengths 27 27 | finished_at stringlengths 27 27 | instruction stringlengths 3.28k 40k | steps stringlengths 3.73k 209k | final_metrics stringlengths 82 91 | verifier_details stringclasses 232
values | attempt int64 0 0 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
construct-ae-p32-spherical-design-l6-s0 | construct-ae-p32-spherical-design | construct | ae | construct-ae-p32-spherical-desig__TkwEiYK | moonshotai/Kimi-K3 | terminus-2 | 0 | AgentTimeoutError | true | 2 | 2 | 2,688 | 6,745 | 2026-10-02T20:34:37.402050Z | 2026-10-02T21:35:01.271680Z | You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.
Format your response as JSON with the following structure:
{
"analysis"... | [{"step_id": 1, "timestamp": "2026-10-02T20:34:47.499989+00:00", "source": "user", "message": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches... | {"total_prompt_tokens": 2688, "total_completion_tokens": 6745, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p32-spherical-design-l2-s0 | construct-ae-p32-spherical-design | construct | ae | construct-ae-p32-spherical-desig__rBDTF7n | moonshotai/Kimi-K3 | terminus-2 | 0 | null | false | 12 | 12 | 161,120 | 20,613 | 2026-10-02T16:32:15.208780Z | 2026-10-02T17:32:38.042780Z | You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.
Format your response as JSON with the following structure:
{
"analysis"... | [{"step_id": 1, "timestamp": "2026-10-02T16:32:24.056795+00:00", "source": "user", "message": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches... | {"total_prompt_tokens": 161120, "total_completion_tokens": 20613, "total_cached_tokens": 0} | {"valid": false, "error": "not a 9-design: some harmonic of degree <= 9 has point average 3.838e-01 > 1e-10", "reward": 0.0} | 0 |
construct-ae-p32-spherical-design-l1-s0 | construct-ae-p32-spherical-design | construct | ae | construct-ae-p32-spherical-desig__uwrBXHa | moonshotai/Kimi-K3 | terminus-2 | 0 | null | false | 12 | 9 | 114,640 | 19,541 | 2026-10-02T17:01:42.451829Z | 2026-10-02T17:26:58.826952Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T17:01:51.455470+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 114640, "total_completion_tokens": 19541, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p38-factorial-factors-l3-s9 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__5yA82oF | moonshotai/Kimi-K3 | terminus-2 | 0 | AgentTimeoutError | true | 8 | 8 | 57,390 | 26,907 | 2026-10-02T20:34:35.560926Z | 2026-10-02T21:35:00.972031Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T20:34:47.084338+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 57390, "total_completion_tokens": 26907, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p38-factorial-factors-l3-s8 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__6LWypEK | moonshotai/Kimi-K3 | terminus-2 | 0 | AgentTimeoutError | true | 5 | 4 | 22,957 | 16,196 | 2026-10-02T16:14:58.290059Z | 2026-10-02T17:15:21.454877Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T16:15:07.094018+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 22957, "total_completion_tokens": 16196, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p38-factorial-factors-l4-s5 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__AQMonyN | moonshotai/Kimi-K3 | terminus-2 | 0 | null | false | 12 | 9 | 170,219 | 38,709 | 2026-10-02T19:33:01.445581Z | 2026-10-02T20:22:14.037519Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T19:33:14.424919+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 170219, "total_completion_tokens": 38709, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p38-factorial-factors-l6-s1 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__EKXT7SK | moonshotai/Kimi-K3 | terminus-2 | 0 | AgentTimeoutError | true | 6 | 5 | 37,763 | 24,512 | 2026-10-02T16:00:26.064971Z | 2026-10-02T17:00:47.170678Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T16:00:33.308929+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 37763, "total_completion_tokens": 24512, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p38-factorial-factors-l2-s2 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__FAdKgqf | moonshotai/Kimi-K3 | terminus-2 | 1 | null | false | 6 | 6 | 32,063 | 10,153 | 2026-10-02T16:33:29.235771Z | 2026-10-02T16:46:46.208442Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T16:33:38.971700+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 32063, "total_completion_tokens": 10153, "total_cached_tokens": 0} | {"valid": true, "min_factor": 28, "reward": 1.0} | 0 |
construct-ae-p38-factorial-factors-l5-s2 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__G6tRdxy | moonshotai/Kimi-K3 | terminus-2 | 0 | AgentTimeoutError | true | 3 | 3 | 8,506 | 15,627 | 2026-10-02T20:34:32.089105Z | 2026-10-02T21:34:59.910274Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T20:34:45.944816+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 8506, "total_completion_tokens": 15627, "total_cached_tokens": 0} | {"valid": false, "error": "no /workdir/answer.json"}
| 0 |
construct-ae-p38-factorial-factors-l2-s7 | construct-ae-p38-factorial-factors | construct | ae | construct-ae-p38-factorial-facto__JW3TAU2 | moonshotai/Kimi-K3 | terminus-2 | 1 | null | false | 12 | 11 | 145,415 | 40,099 | 2026-10-02T20:34:43.221484Z | 2026-10-02T21:23:19.665307Z | "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be (...TRUNCATED) | "[{\"step_id\": 1, \"timestamp\": \"2026-10-02T20:34:52.338726+00:00\", \"source\": \"user\", \"mess(...TRUNCATED) | {"total_prompt_tokens": 145415, "total_completion_tokens": 40099, "total_cached_tokens": 0} | {"valid": true, "min_factor": 39, "reward": 1.0} | 0 |
End of preview. Expand in Data Studio
rlmath Kimi-K3 agent trajectories
Agent trajectories of moonshotai/Kimi-K3 solving tasks from rlmath. rlmath is a collection of math construction and optimization environments with deterministic programmatic verifiers and no LLM judge, packaged as Harbor tasks. The agent is Harbor's Terminus-2. It works in a sandboxed terminal with no network: it writes and runs code, then submits a construction that the task's grader scores.
This is a partial collection. 1519 trajectories over 1519 tasks. The planned run is pass@2 over 3,456 training tasks (6,912 trials). It paused when the Kimi-K3 endpoint went offline. Harbor runs tasks in alphabetical order, so this snapshot is weighted toward construct-* tasks.
Collection setup
- Model: Kimi-K3, served with SGLang and reached through an OpenAI-compatible endpoint. Reasoning is returned separately and stored per step as
reasoning_content. - Agent: Harbor 0.23.0
terminus-2, at most 12 turns,max_tokens32,768 per call, no context summarization. Agent timeout is 3,600 s. - Tasks: the rlmath training split, which excludes 121 held-out validation instances. Sandboxes are Docker with no network.
- Reward:
construct-*tasks score 1 or 0.optimize-*tasks score 0 for an invalid submission, otherwise 0.1 + 0.9 · clipped progress toward the target. - Excluded: trials that failed only because the endpoint had no provider. Trials that hit the agent timeout are kept (
agent_timed_out=True); they were scored on whatever the agent had produced.
Fields
| field | meaning |
|---|---|
task_name, family, task_type, source_suite, attempt |
task instance, family (instance suffix stripped), construct/optimize, source suite prefix, attempt index |
reward |
verifier reward |
exception_type, agent_timed_out |
Harbor exception, if any |
n_agent_steps, n_steps_with_reasoning, total_prompt_tokens, total_completion_tokens |
trajectory statistics |
instruction |
the task prompt shown to the agent |
steps |
JSON list of ATIF steps (user, agent and environment messages, reasoning_content, tool calls, observations, metrics) |
verifier_details |
the grader's details.json |
final_metrics |
agent-level metrics |
raw_harbor_job_k3train.tar.gz holds the raw Harbor job directory: per-trial configs, logs, verifier output and trajectories.
Snapshot statistics
Mean reward 0.959 over 1518 scored trajectories. By type: construct: 1480 trials, mean 0.972, optimize: 38 trials, mean 0.464.
- Downloads last month
- 27