datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
agentic-reasoning-trace-summaries-40k
Reasoning Summary JSON Dataset
We built this dataset to train models to turn long reasoning/work traces into short structured summaries.
Each example has a verbose trace in input and a compact JSON summary in output. The summary is shaped like the kind of progress update we want a model to produce while it is working: a title, a subtitle, a short summary, and the current task.
The dataset is JSONL with 43,734 rows. The rows are ordered from longest to shortest so long-context… See the full description on the dataset page: https://huggingface.co/datasets/Qyrou/agentic-reasoning-trace-summaries-40k.tw-agentic-reasoning-instruct
tw-agentic-reasoning-instruct
繁體中文(台灣)的代理式推理軌跡。跟一般指令資料集不同,這裡每一筆不是
「請模型寫一篇答案」,而是讓一個 agent 真的跑一輪:它會去查網路、呼叫
工具、拿到真實回傳、踩到工具的限制,然後才給出答案。工具結果是真的查回來
的,模型講不出來的時候也是真的講不出來。
種子題目取自 twinkle-ai/tw-reasoning-instruct-50k
的 input 欄位;軌跡由 ai-twinkle/harness
驅動 Qwen3.8-27B 產生,搜尋能力來自自架的 SearXNG(透過 MCP)。
這是進行中的版本。 生成尚未跑完,之後會更新為完整版,並加入「帶繁中
系統提示」的第二組條件。
格式
messages 是 OpenAI 的形狀(因為要表達 tool_calls),推理依照
tw-reasoning-instruct-50k 的慣例包在 <think> 標籤裡:
<think>
(模型的推理)
</think>
(給使用者的答案)… See the full description on the dataset page: https://huggingface.co/datasets/lianghsun/tw-agentic-reasoning-instruct.
