Datasets:
BFCL-toolcall-SDG
資料集概述
- 來源:依 Berkeley Function Calling Leaderboard(BFCL) 的評測類別設計的合成工具呼叫對話,經人工審查。
- 用途:工具呼叫(function calling)的監督式微調(SFT)。
- 主要輸出:
data/train.jsonl(373 列、366 條對話軌跡)。 - 授權:僅供內部共享使用。
Splits
| Split | 檔案 | 說明 |
|---|---|---|
train |
data/train.jsonl |
單輪與多輪工具呼叫對話,每列是一個訓練樣本 |
資料結構
- 每一列的
messages都以一則 assistant 訊息結尾,這則訊息就是監督目標,內容可能是工具呼叫(154 列),也可能是文字回覆(219 列)。 - 單輪類別(
simple、multiple、parallel、live_*、irrelevance等):每條軌跡一列。 - 多輪類別(
multi_turn_*):每條軌跡一列,內容是包含完整對話的最後一步。
欄位說明
| 欄位 | 型別 | 說明 |
|---|---|---|
id |
string | 每列唯一的識別碼 |
trajectory_id |
string | 所屬對話軌跡;只有 multi_turn_miss_func 會有多列共用同一個值 |
category |
string | 對應的 BFCL 類別(見下方統計) |
messages |
list | OpenAI 格式的對話:user、assistant(可帶 tool_calls)、tool(帶 tool_call_id、name);沒有 system 訊息 |
tools |
string | 工具定義清單的 JSON 字串;json.loads 後為 OpenAI function 格式的 list({"type": "function", "function": {...}}) |
tool_calls[].function.arguments 與 OpenAI 格式相同,存成 JSON 字串。
目前資料統計
data/train.jsonl:373 列,其中單輪 305 列(305 條軌跡)、多輪 68 列(61 條軌跡)。
| category | 軌跡數 | 列數 |
|---|---|---|
simple |
126 | 126 |
multiple |
59 | 59 |
parallel |
39 | 39 |
multi_turn_base |
28 | 28 |
multi_turn_miss_param |
18 | 18 |
parallel_multiple |
18 | 18 |
irrelevance |
14 | 14 |
live_parallel_multiple |
14 | 14 |
multi_turn_long_context |
14 | 14 |
live_relevance |
11 | 11 |
live_multiple |
9 | 9 |
multi_turn_miss_func |
1 | 8 |
live_irrelevance |
7 | 7 |
live_parallel |
5 | 5 |
live_simple |
3 | 3 |
- 語言:單輪類別為繁體中文(台灣用語);多輪類別以英文為主(約九成),其餘為繁體中文;工具描述多為英文。
使用方式(最小範例)
import json
from datasets import load_dataset
ds = load_dataset("minyichen/BFCL-toolcall-SDG", split="train")
row = ds[0]
tools = json.loads(row["tools"]) # 工具定義
for m in row["messages"]:
for tc in m.get("tool_calls") or []:
args = json.loads(tc["function"]["arguments"]) # 呼叫參數
若要直接讀檔:
import json
with open("data/train.jsonl", encoding="utf-8") as f:
rows = [json.loads(line) for line in f]
- Downloads last month
- 48