Datasets:

ZHIYII
/

AgentBrew_Github

messages listlengths 3 3	weight float64 0 2	task_id stringlengths 21 23	step_id int64 0 50
[ { "content": "You are an agent for repository management.", "role": "system" }, { "content": "You are a ReAct (Reasoning and Acting) agent.\nYou are an agent for repository management.\n\n### Tools Description ###\\n\\nServer: github\\nTool: add_comment_to_pending_review\\nDescription:\\nAdd review ...	0.766282	00_github_task_0001_12	0
[ { "content": "You are an agent for repository management.", "role": "system" }, { "content": "You are a ReAct (Reasoning and Acting) agent.\nYou are an agent for repository management.\n\n### Tools Description ###\\n\\nServer: github\\nTool: add_comment_to_pending_review\\nDescription:\\nAdd review ...	2	00_github_task_0001_12	2
[ { "content": "You are an agent for repository management.", "role": "system" }, { "content": "You are a ReAct (Reasoning and Acting) agent.\nYou are an agent for repository management.\n\n### Tools Description ###\\n\\nServer: github\\nTool: add_comment_to_pending_review\\nDescription:\\nAdd review ...	0.655833	00_github_task_0001_12	3
[ { "content": "You are an agent for repository management.", "role": "system" }, { "content": "You are a ReAct (Reasoning and Acting) agent.\nYou are an agent for repository management.\n\n### Tools Description ###\\n\\nServer: github\\nTool: add_comment_to_pending_review\\nDescription:\\nAdd review ...	0.754094	00_github_task_0001_12	4
[{"content":"You are an agent for repository management.","role":"system"},{"content":"You are a ReA(...TRUNCATED)	0.337438	00_github_task_0001_12	5
[{"content":"You are an agent for repository management.","role":"system"},{"content":"You are a ReA(...TRUNCATED)	1.322331	00_github_task_0001_12	6
[{"content":"You are an agent for repository management.","role":"system"},{"content":"You are a ReA(...TRUNCATED)	1	00_github_task_0001_12	7
[{"content":"You are an agent for repository management.","role":"system"},{"content":"You are a ReA(...TRUNCATED)	0.770128	00_github_task_0001_16	0
[{"content":"You are an agent for repository management.","role":"system"},{"content":"You are a ReA(...TRUNCATED)	2	00_github_task_0001_16	2
[{"content":"You are an agent for repository management.","role":"system"},{"content":"You are a ReA(...TRUNCATED)	0.421626	00_github_task_0001_16	3

End of preview. Expand in Data Studio

AgentBrew GitHub

AgentBrew GitHub is a public weighted action-level SFT dataset for training GitHub tool-use agents. It is converted from ZHIYII/Github_Entropy_SFT_swift for the AgentBrew weighted-SFT training pipeline.

Processing

Parsed messages from a serialized JSON string into a native list of chat messages.
Kept only rows with a finite weight > 0.
Removed is_final_answer, is_error_step, and raw_advantage.
Retained task_id and step_id for provenance and analysis.
Preserved the source dataset's positive training weights without further normalization.

Schema

Field	Type	Description
`messages`	list of objects	`system`, `user`, and `assistant` chat messages.
`weight`	float	Positive per-sample weighted-SFT multiplier.
`task_id`	string	Source task identifier.
`step_id`	integer	Action index within the source task.

Statistics

Training rows: 25,993
Unique tasks: 3,426
Weight range: 0.0003986843417 to 2
Mean weight: 0.9989055967
Role pattern: system -> user -> assistant

Loading

from datasets import load_dataset

dataset = load_dataset("ZHIYII/AgentBrew_Github", split="train")

The training objective averages causal-language-model loss over valid assistant tokens for each sample, multiplies it by weight, and then averages the weighted sample losses across the batch.

Downloads last month: 60