Datasets:

metaevo-ai
/

harness-zero-sft

task_id stringlengths 9 33	trial stringlengths 18 29	turns int64 2 48	masked_reasoning_turns listlengths 0 3	replacements int64 0 5	kind stringclasses 1 value	messages listlengths 5 119	masked_assistant_turns listlengths 10 10 ⌀
uspto-train-278	uspto-train-278__bngdRFb__env	6	[]	5	session	[ { "role": "system", "content": "You are an agent that solves tasks in a Linux sandbox.\n\nYou have one tool, `execute`, which runs one bash command. Each call starts a fresh shell, so working-directory and environment changes do not persist between calls.\n\nWork on the task by issuing one `execute` call at...	null
spreadsheetbench-verified__534-26	534-26__PcjoJpU__env	6	[]	2	session	[ { "role": "system", "content": "You are an agent that solves tasks in a Linux sandbox.\n\nYou have one tool, `execute`, which runs one bash command. Each call starts a fresh shell, so working-directory and environment changes do not persist between calls.\n\nWork on the task by issuing one `execute` call at...	null
uspto-train-173	uspto-train-173__FRdvFsY__env	4	[]	4	session	[ { "role": "system", "content": "You are an agent that solves tasks in a Linux sandbox.\n\nYou have one tool, `execute`, which runs one bash command. Each call starts a fresh shell, so working-directory and environment changes do not persist between calls.\n\nWork on the task by issuing one `execute` call at...	null
spreadsheetbench-verified__51262	51262__8ef9CuW__env	4	[ 1 ]	3	session	[ { "role": "system", "content": "You are an agent that solves tasks in a Linux sandbox.\n\nYou have one tool, `execute`, which runs one bash command. Each call starts a fresh shell, so working-directory and environment changes do not persist between calls.\n\nWork on the task by issuing one `execute` call at...	null
spreadsheetbench-verified__382-29	382-29__e6pMoMR__env	5	[]	3	session	[ { "role": "system", "content": "You are an agent that solves tasks in a Linux sandbox.\n\nYou have one tool, `execute`, which runs one bash command. Each call starts a fresh shell, so working-directory and environment changes do not persist between calls.\n\nWork on the task by issuing one `execute` call at...	null
spreadsheetbench-verified__54638	54638__h53WJan__env	25	[ 3 ]	3	session	[ { "role": "system", "content": "You are an agent that solves tasks in a Linux sandbox.\n\nYou have one tool, `execute`, which runs one bash command. Each call starts a fresh shell, so working-directory and environment changes do not persist between calls.\n\nWork on the task by issuing one `execute` call at...	null
spreadsheetbench-verified__47933	47933__AgJscMY__env	12	[]	5	session	[{"role":"system","content":"You are an agent that solves tasks in a Linux sandbox.\n\nYou have one (...TRUNCATED)	null
spreadsheetbench-verified__560-12	560-12__5qZtgE6__env	6	[]	5	session	[{"role":"system","content":"You are an agent that solves tasks in a Linux sandbox.\n\nYou have one (...TRUNCATED)	null
spreadsheetbench-verified__382-10	382-10__roRZbyb__env	5	[]	3	session	[{"role":"system","content":"You are an agent that solves tasks in a Linux sandbox.\n\nYou have one (...TRUNCATED)	null
spreadsheetbench-verified__10452	10452__HNw55gX__env	4	[]	2	session	[{"role":"system","content":"You are an agent that solves tasks in a Linux sandbox.\n\nYou have one (...TRUNCATED)	null

End of preview. Expand in Data Studio

Harness-Zero SFT trajectories

Reviewed training trajectories for Harness-Zero (harness distillation via agent-as-harness), used to LoRA-SFT Qwen3.5-9B into the released checkpoints (ahd-9b-ssb, ahd-9b-appworld, ahd-9b-uspto).

sessions.jsonl — 1,269 reviewed rollouts, one JSON object per line: {task_id, trial, turns, replacements, kind, messages, masked_reasoning_turns[, masked_assistant_turns]}. messages is the student-visible trajectory only ({role, content, tool_calls?, tool_call_id?}), where content is a string or a list of thinking / text blocks; teacher review exchanges never appear. Turns listed in masked_reasoning_turns / masked_assistant_turns carry reviewer-perspective reasoning or demonstration text and are masked out of the loss.

Statistics

Source	Sessions	Unique tasks
SpreadsheetBench Verified (300-task train split)	487	264
AppWorld (147-task train split)	282	133
USPTO retrosynthesis (500-task train split)	500	500
Total	1,269	897

data_sha256 = 31e1b9424cdc1604a85834169edce408a66706655ef5505a0768ee7d313de3e4

Notes

AppWorld sessions embed the official AppWorld ReAct demonstration in their teaching prefix (masked from the loss); redistribution follows the upstream AppWorld license.
Credentials visible inside AppWorld trajectories (app passwords, JWTs) are synthetic per-task sandbox credentials generated by the AppWorld simulator, as in any AppWorld trajectory — they are not real secrets.
Code: Harness-Zero repository (data collection, datum building, and the Tinker training recipe that consumes this file).

Downloads last month: 42

Harness-Zero SFT trajectories

Contents

Statistics

Notes