SoMu1994/MXMUHPXTWS

Multi-source SFT dataset combining several subsets, each tagged via the source column:

source	description
alltopics	broad multi-topic conversations
terminal	Terminal/agentic traces
opencode_focus_v8	OpenCode-style focused traces v8
opencode_focus_v9	OpenCode-style focused traces v9
opencode_focus_v10	OpenCode-style focused traces v10
opencode_focus_v11	OpenCode-style focused traces v11

Schema

All rows share a uniform schema. Missing string fields are normalized to "" and missing list fields are normalized to [] so the parquet file has a strict single schema across all subsets.

messages: list<struct<
    role: string,
    content: string,                 # JSON-encoded if originally a list (e.g. multimodal)
    reasoning_content: string,       # "" if absent
    refusal: string,                 # "" if absent
    tool_call_id: string,            # "" if absent
    tool_calls: list<struct<
        id: string,
        type: string,
        function: struct<name: string, arguments: string>   # arguments is a JSON string
    >>                               # [] if absent
>>
tools: list<struct<
    type: string,
    function: struct<
        name: string,
        description: string,
        parameters: string           # JSON-encoded JSON schema
    >
>>
tool_choice: string
timeout: bool
is_subagent: bool
truncated: bool
source: string

Loading

from datasets import load_dataset
ds = load_dataset("SoMu1994/MXMUHPXTWS", split="train")
alltopics = ds.filter(lambda r: r["source"] == "alltopics")

Tool-call arguments and tool parameter schemas are stored as JSON-encoded strings; decode them with json.loads to recover their structure.