Datasets:

PatronusAI
/

few-word-do-trick-traces

model stringclasses 14 values	model_key stringclasses 14 values	model_id stringclasses 14 values	benchmark stringclasses 1 value	task stringclasses 8 values	reward float64 1 1	steps listlengths 5 64	final_metrics unknown
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	django__django-11099	1	[ { "step_id": 1, "timestamp": "2026-11-06T04:11:09.423252+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 ...	{ "total_prompt_tokens": 14110, "total_completion_tokens": 1513, "total_cached_tokens": 5568, "total_cost_usd": 0.027010444300000002 }
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	django__django-11133	1	[ { "step_id": 1, "timestamp": "2026-11-06T04:11:20.805850+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 ...	{ "total_prompt_tokens": 63754, "total_completion_tokens": 4687, "total_cached_tokens": 44480, "total_cost_usd": 0.0941248 }
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	django__django-11179	1	[ { "step_id": 1, "timestamp": "2026-11-06T04:11:24.499515+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 ...	{ "total_prompt_tokens": 27704, "total_completion_tokens": 2581, "total_cached_tokens": 8768, "total_cost_usd": 0.0649888 }
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	django__django-14089	1	[ { "step_id": 1, "timestamp": "2026-11-06T04:13:05.324577+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 ...	{ "total_prompt_tokens": 26679, "total_completion_tokens": 1786, "total_cached_tokens": 7424, "total_cost_usd": 0.05499408 }
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	django__django-16139	1	[{"step_id":1,"timestamp":"2026-11-06T04:12:41.627711+00:00","source":"user","message":"You are an A(...TRUNCATED)	{"total_prompt_tokens":47101,"total_completion_tokens":3731,"total_cached_tokens":22592,"total_cost_(...TRUNCATED)
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	scikit-learn__scikit-learn-13439	1	[{"step_id":1,"timestamp":"2026-11-06T04:12:58.666534+00:00","source":"user","message":"You are an A(...TRUNCATED)	{"total_prompt_tokens":25894,"total_completion_tokens":2127,"total_cached_tokens":10944,"total_cost_(...TRUNCATED)
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	astropy__astropy-14995	1	[{"step_id":1,"timestamp":"2026-11-06T04:13:50.036403+00:00","source":"user","message":"You are an A(...TRUNCATED)	{"total_prompt_tokens":54375,"total_completion_tokens":3894,"total_cached_tokens":23680,"total_cost_(...TRUNCATED)
Kimi K3.1 Pro	kimi31	openrouter/moonshotai/kimi-k3.1-pro	swe-bench-verified	pytest-dev__pytest-7432	1	[{"step_id":1,"timestamp":"2026-11-06T04:13:58.284795+00:00","source":"user","message":"You are an A(...TRUNCATED)	{"total_prompt_tokens":59174,"total_completion_tokens":3702,"total_cached_tokens":32448,"total_cost_(...TRUNCATED)
Kimi K3	kimi	openrouter/moonshotai/kimi-k3	swe-bench-verified	django__django-11099	1	[{"step_id":1,"timestamp":"2026-11-06T04:03:10.451758+00:00","source":"user","message":"You are an A(...TRUNCATED)	{"total_prompt_tokens":8676,"total_completion_tokens":1629,"total_cached_tokens":2688,"total_cost_us(...TRUNCATED)
Kimi K3	kimi	openrouter/moonshotai/kimi-k3	swe-bench-verified	django__django-11133	1	[{"step_id":1,"timestamp":"2026-11-06T04:03:24.569540+00:00","source":"user","message":"You are an A(...TRUNCATED)	{"total_prompt_tokens":96852,"total_completion_tokens":4894,"total_cached_tokens":66752,"total_cost_(...TRUNCATED)

End of preview. Expand in Data Studio

Few Word Do Trick traces

Agent trajectories behind the Patronus AI write-up "Few Word Do Trick: Open Models Cut the Filler from Their Reasoning" (Arnhav Datar, Varun Gangal, November 2026).

Each row is one run of one open model on one agentic coding task, with the full step-by-step trajectory, including each step's reasoning text, commands and observations.

Split	Tasks	Runs
`swe_bench_verified`	8 SWE-bench Verified tasks	112
`swe_bench_pro`	5 SWE-bench Pro tasks	60
`terminal_bench_2_1`	7 Terminal-Bench 2.1 tasks	84

Twelve models ran all 20 tasks: Kimi K3.1 Pro, Kimi K3, GLM-5.3-Flash, GLM-5.2, DeepSeek V4.1 Flash (1017), DeepSeek V4 Flash, Reflection Beam 1.0, MiniMax-M3.1-Flash, Inkling 2 Small, Muse Spark Open, MiMo V2.6 Flash and Qwen3.8 Flash. Two more ran only the SWE-bench Verified tasks: DeepSeek V4 Flash (0731) and GLM-5.3.

Fields

model: display name used in the write-up.
model_key: short key used in the analysis code.
model_id: the provider model ID the agent called.
benchmark, task: where the run comes from.
reward: 1.0 if the task's tests passed, 0.0 if not.
steps: the agent trajectory (ATIF format). Agent steps carry reasoning_content (the model's reasoning), message, tool_calls and observation.
final_metrics: token and cost totals reported by the harness.

Setup

Harbor with the Terminus-2 agent, one run per model per task, reasoning effort set to high (0.99 on Tinker for Inkling 2 Small), temperature 1.0, 262k-token context cap. Reflection Beam 1.0 was run with its earlier reasoning passed back on each turn, because its providers require it; the other models did not see their earlier reasoning.

Downloads last month: 10

Few Word Do Trick traces

Contents

Fields

Setup