The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Error code: DatasetGenerationError
Exception: ValueError
Message: Value is too big!
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 281, in _generate_tables
examples = [ujson_loads(line) for line in batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Value is too big!
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
id string | messages list | tools list | meta dict |
|---|---|---|---|
nemotron_tool_calling-000002 | [
{
"role": "user",
"content": "I've been wanting to catch the landscape photography exhibit at the Portland Art Museum later this month, but it's been a while since I've been there. Before planning a trip, could you check if there are any suitable dining options nearby that accept reservations? Also, it woul... | [
{
"type": "function",
"function": {
"name": "verify_schedule",
"description": "Verify if the time and location of the specified schedule are correct",
"parameters": {
"type": "object",
"properties": {
"schedule_id": {
"type": "string",
"des... | {
"source": "nemotron_tool_calling",
"orig_uuid": "cb41c97a-14aa-451e-a628-13732f275b4b",
"reasoning_turns": 2,
"calls_without_reasoning": 0,
"score": 2.5917,
"n_user_turns": 1,
"n_call_turns": 2,
"n_calls": 2,
"max_parallel": 1,
"n_tools": 15,
"domains": [
"productivity"
],
"tokens": 2801... |
nemotron_tool_calling-000003 | [
{
"role": "user",
"content": "Hey, I’ve been thinking about moving up from courier work into a logistics or dispatch role here in St. Pete. Can you help me find job openings that fit my experience?"
},
{
"role": "assistant",
"content": "",
"tool_calls": [
{
"id": "call_1",
... | [
{
"type": "function",
"function": {
"name": "download_attachment",
"description": "Download attachments from a specified email account",
"parameters": {
"type": "object",
"properties": {
"email": {
"type": "string",
"description": "Email ad... | {
"source": "nemotron_tool_calling",
"orig_uuid": "a467f744-f011-4f78-be6c-f27a76b78248",
"reasoning_turns": 2,
"calls_without_reasoning": 0,
"score": 2.6667,
"n_user_turns": 1,
"n_call_turns": 2,
"n_calls": 2,
"max_parallel": 1,
"n_tools": 15,
"domains": [
"banking",
"websearch"
],
"t... |
nemotron_tool_calling-000004 | [
{
"role": "user",
"content": "My stove has been acting up lately - the burners keep clicking but won't stay lit. I need to find a reliable appliance repair service here in Hendersonville, compare at least three different companies based on customer reviews, verify their credentials, then schedule the earlie... | [
{
"type": "function",
"function": {
"name": "appliance_repair_cancellation",
"description": "Appliance Repair Cancellation Tool",
"parameters": {
"type": "object",
"properties": {
"requestID": {
"description": "The repair request number to be cancelled... | {
"source": "nemotron_tool_calling",
"orig_uuid": "9cbe370f-d27b-43a3-8a50-bf58a66cc566",
"reasoning_turns": 3,
"calls_without_reasoning": 0,
"score": 3.4583,
"n_user_turns": 1,
"n_call_turns": 3,
"n_calls": 3,
"max_parallel": 1,
"n_tools": 15,
"domains": [
"messaging",
"productivity"
],... |
nemotron_tool_calling-000005 | [
{
"role": "user",
"content": "I'm planning a system maintenance task focused on servers located in multiple cities. How can I navigate to a specific city and then execute a system health check command on the local servers once I've reached the destination?"
},
{
"role": "assistant",
"content": "... | [
{
"type": "function",
"function": {
"name": "navigate_to",
"description": "navigate to the city.\n\n Args:\n city: destination city name",
"parameters": {
"type": "object",
"properties": {
"city": {
"title": "City",
"type": "strin... | {
"source": "nemotron_tool_calling",
"orig_uuid": "7dd45783-a374-4e64-9914-ac0dc9fcd90c",
"reasoning_turns": 8,
"calls_without_reasoning": 0,
"score": 4.6833,
"n_user_turns": 1,
"n_call_turns": 8,
"n_calls": 8,
"max_parallel": 1,
"n_tools": 10,
"domains": [
"navigation"
],
"tokens": 3342
} |
nemotron_tool_calling-000006 | [
{
"role": "user",
"content": "We need to launch a globally‑available feature that automatically displays the current trending Lottie animation styles on our web app. The solution must retrieve the top‑ranked animation styles and then provision Azure resource groups in East US, West Europe, and South Asia th... | [
{
"type": "function",
"function": {
"name": "search_animations",
"description": "Search for Lottie animations by keywords, tags, and other criteria. Supports pagination.",
"parameters": {
"type": "object",
"properties": {
"query": {
"type": "string",
... | {
"source": "nemotron_tool_calling",
"orig_uuid": "9060c72a-53d0-43de-8315-053c412bab9b",
"reasoning_turns": 3,
"calls_without_reasoning": 0,
"score": 2.8833,
"n_user_turns": 1,
"n_call_turns": 3,
"n_calls": 3,
"max_parallel": 1,
"n_tools": 10,
"domains": [],
"tokens": 11792
} |
nemotron_tool_calling-000007 | [
{
"role": "user",
"content": "I just noticed Bitcoin's price suddenly dropped 8% in the last hour and there's unusual volume on the derivatives markets. What's happening right now in the crypto space that might be causing this sudden market movement?"
},
{
"role": "assistant",
"content": "",
... | [
{
"type": "function",
"function": {
"name": "getNews",
"description": "Get the latest news flash about crypto (usually the up-to-date events). It can be used to query the latest happened event in crypto community. (e.g. the famous CEX launched a new activity, a trending project emerges in market... | {
"source": "nemotron_tool_calling",
"orig_uuid": "13f946b7-5368-44a2-ae0b-73db0d9d6ead",
"reasoning_turns": 2,
"calls_without_reasoning": 0,
"score": 2.7117,
"n_user_turns": 1,
"n_call_turns": 2,
"n_calls": 2,
"max_parallel": 1,
"n_tools": 2,
"domains": [
"websearch"
],
"tokens": 5775
} |
nemotron_tool_calling-000008 | [
{
"role": "user",
"content": "Could you locate the newest arXiv paper on AI alignment and give me its title, abstract, authors, and PDF link?"
},
{
"role": "assistant",
"content": "",
"tool_calls": [
{
"id": "call_1",
"type": "function",
"function": {
... | [
{
"type": "function",
"function": {
"name": "scrape_recent_category_papers",
"description": "[크롤링] 특정 카테고리의 'recent' 페이지를 스크랩하여 최신 논문 목록을 가져옵니다.",
"parameters": {
"type": "object",
"properties": {
"category": {
"title": "Category",
"type": ... | {
"source": "nemotron_tool_calling",
"orig_uuid": "e19797ec-71c6-410c-834b-a3a18948b998",
"reasoning_turns": 6,
"calls_without_reasoning": 0,
"score": 5.2033,
"n_user_turns": 4,
"n_call_turns": 6,
"n_calls": 6,
"max_parallel": 1,
"n_tools": 7,
"domains": [],
"tokens": 6990
} |
nemotron_tool_calling-000010 | [
{
"role": "user",
"content": "Could you search for any upcoming photography exhibits at the New Haven Museum of Art or nearby galleries? Also, I’d like to know what the weather will be like in Hamden for the next weekend, in case I decide to hike the Sleeping Giant Trail or catch a cricket match."
},
{
... | [
{
"type": "function",
"function": {
"name": "bookhotel",
"description": "Book a hotel API that takes multiple parameters such as 'hotel_name', 'check_in_date', 'check_out_date', 'room_count', 'adult_count', 'child_count'. The 'check_in_date' parameter is the date when the guest is supposed to ar... | {
"source": "nemotron_tool_calling",
"orig_uuid": "da253e9f-5646-4773-8f0a-7aef1f139f83",
"reasoning_turns": 2,
"calls_without_reasoning": 0,
"score": 2.9917,
"n_user_turns": 1,
"n_call_turns": 2,
"n_calls": 2,
"max_parallel": 1,
"n_tools": 15,
"domains": [
"airline",
"navigation",
"tr... |
nemotron_tool_calling-000011 | [
{
"role": "user",
"content": "I’d like to determine the average points scored per team from the latest NBA season summary. Could you first retrieve the summary’s metadata, which includes the reported total points for each team, and then calculate the average by dividing the combined points across all teams ... | [
{
"type": "function",
"function": {
"name": "add",
"description": "Add two numbers",
"parameters": {
"type": "object",
"properties": {
"a": {
"title": "A",
"type": "integer"
},
"b": {
"title": "B",
... | {
"source": "nemotron_tool_calling",
"orig_uuid": "69ea4ca4-becb-4dac-b282-9246d28a5fc1",
"reasoning_turns": 2,
"calls_without_reasoning": 0,
"score": 2.9217,
"n_user_turns": 1,
"n_call_turns": 2,
"n_calls": 2,
"max_parallel": 1,
"n_tools": 8,
"domains": [
"math"
],
"tokens": 3901
} |
nemotron_tool_calling-000012 | [
{
"role": "user",
"content": "I'm helping organize a senior citizen bus trip from Sayreville, NJ to see a Yankees game at Yankee Stadium next month, and I need to work out the logistics. First, could you find the exact postal code for Yankee Stadium in the Bronx? Then I'll need you to plan the best driving ... | [
{
"type": "function",
"function": {
"name": "ai_search_tool",
"description": "An intelligent search tool capable of fuzzy and deep search",
"parameters": {
"type": "object",
"properties": {
"query": {
"description": "Search keywords, must be between 2 ... | {
"source": "nemotron_tool_calling",
"orig_uuid": "dc034a7b-9211-46b6-ada0-c855ed1a8f1c",
"reasoning_turns": 4,
"calls_without_reasoning": 0,
"score": 4.125,
"n_user_turns": 1,
"n_call_turns": 4,
"n_calls": 4,
"max_parallel": 1,
"n_tools": 15,
"domains": [
"navigation",
"websearch"
],
... |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
final_train_data
Dữ liệu SFT cho function calling dành cho fc_baselines, được
lọc từ hai bộ dữ liệu huấn luyện trong train_data/ (NVIDIA
Nemotron-SFT-Agentic-v2 và Nanbeige ToolMind) và chuyển về định dạng bản ghi
trong README của fc_baselines.
Hai biến thể từ cùng một lần lựa chọn được phát hành, mỗi biến thể ứng với một
reasoning_mode:
final_train_data/
├── no_reasoning/ # mặc định của configs/*/train_*.yaml
│ ├── train.jsonl 97,315 bản ghi (1.75 GB)
│ └── validation.jsonl 1,979 bản ghi
├── reasoning/ # giữ reasoning_content ở các lượt gọi công cụ
│ ├── train.jsonl 97,203 bản ghi (1.75 GB)
│ └── validation.jsonl 1,999 bản ghi
├── stats.json # số lượng theo nguồn do build ghi ra
└── README.md
fc_baselines nhận một file .jsonl hoặc một thư mục chỉ chứa các file *.jsonl,
nên config có thể trỏ thẳng đến các file này sau khi dataset được tải về hoặc mount.
Các khóa cấu hình phải dùng là
train_data và eval_data:
train_data: path/to/no_reasoning/train.jsonl
eval_data: path/to/no_reasoning/validation.jsonl
reasoning_mode: no_reasoning # phải khớp với thư mục đã chọn
cutoff_len: 16384
Mỗi bản ghi là một đối tượng JSON duy nhất, trong đó hội thoại và các công cụ được giữ tách biệt:
{
"id": "toolmind_graphsyn-000042",
"messages": [
{"role": "user", "content": "What's the current time in New York?"},
{"role": "assistant", "content": "", "tool_calls": [
{"id": "call_1", "type": "function",
"function": {"name": "get_current_time", "arguments": {"timezone": "America/New_York"}}}]},
{"role": "tool", "tool_call_id": "call_1", "content": "{\"time\": \"10:00 AM\"}"},
{"role": "assistant", "content": "The current time in New York is 10:00 AM."}
],
"tools": [
{"type": "function", "function": {
"name": "get_current_time",
"description": "Get the current time",
"parameters": {"type": "object",
"properties": {"timezone": {"type": "string", "description": "…"}},
"required": ["timezone"]}}}
],
"meta": {"source": "toolmind_graphsyn", "score": 5.31, "n_call_turns": 2, "tokens": 481, "…": "…"}
}
id và meta bị bỏ qua khi huấn luyện. meta chứa file nguồn, điểm liên quan,
các số liệu hình dạng được dùng để chọn bản ghi và độ dài token chính xác.
Các id được dùng chung giữa hai biến thể. Một bản ghi tồn tại ở cả hai chế độ sẽ
có cùng id trong no_reasoning/ và reasoning/, đồng thời các bản ghi giữ lại để
đánh giá cũng giống nhau, nên có thể so sánh hai lần chạy theo từng bản ghi.
Nội dung được chọn và lý do
Ba benchmark trong test_data/ — CAR-bench, tau^3-bench và BFCL V4 — kiểm thử
hành vi nhiều hơn chủ đề: sử dụng công cụ qua nhiều bước, các lệnh gọi sử dụng kết
quả trước đó, lệnh gọi song song, hỏi thay vì đoán, từ chối những việc không công cụ
nào thực hiện được và tuân theo chính sách được viết trong một hội thoại dài. Bộ lọc
chấm điểm mọi ứng viên đúng theo các đặc tính đó (scripts/score_records.py), cộng
thêm một mức nhỏ cho các miền mà benchmark đề cập đến, rồi lấy quota từng nguồn từ
trên xuống dưới với giới hạn theo từng nhóm đa dạng, để không một môi trường hay
dạng hội thoại nào chiếm hết quota.
| Nguồn | Kho dữ liệu | không suy luận | có suy luận | Lý do đưa vào |
|---|---|---|---|---|
nemotron_tool_calling |
655,686 | 36,658 | 47,066 | sử dụng công cụ qua nhiều bước trên nhiều bộ công cụ |
nemotron_customer_service |
278,855 | 24,000 | 24,000 | hội thoại có chính sách trong system prompt, đúng dạng tau^3 |
toolmind_graphsyn |
155,940 | 13,983 | 13,995 | đi qua toàn bộ đồ thị công cụ: các lệnh gọi nối tiếp |
toolmind_apigen_mt |
24,734 | 6,999 | 1,438 | quỹ đạo nhiều lượt dài với kết quả thực |
toolmind_tau_train |
10,618 | 5,000 | 48 | split retail train của tau-bench |
toolmind_button |
20,919 | 3,999 | 4,000 | phân rã thành các lệnh gọi tuần tự |
toolmind_when2call |
17,515 | 3,500 | 3,500 | khi không nên gọi công cụ và khi nào nên hỏi |
toolmind_toolace |
7,314 | 2,500 | 2,500 | các bề mặt API rộng và khác thường |
toolmind_glaive |
20,010 | 1,000 | 1,000 | hành vi nền tảng với một lệnh gọi đơn giản |
toolmind_xlam |
100,868 | 1,000 | 1,000 | độ bao phủ lệnh gọi song song |
nemotron_search |
655 | 655 | 655 | quỹ đạo tìm kiếm web dài |
| Tổng | 1,293,114 | 99,294 | 99,202 |
Thành phần của tập được chọn:
| không suy luận | có suy luận | |
|---|---|---|
| token (template Qwen3.5) | 477M, trung vị 4,456, p95 10,424 | 446M, trung vị 4,097, p95 10,467 |
| 4+ lượt gọi công cụ | 41,618 (42%) | 23,997 (24%) |
| 2–3 lượt gọi công cụ | 39,095 (39%) | 31,793 (32%) |
| 1 lượt gọi công cụ | 15,949 (16%) | 38,442 (39%) |
| không gọi công cụ (hỏi / từ chối) | 2,632 (3%) | 4,970 (5%) |
| một lượt có các lệnh gọi song song | 14,275 (14%) | 14,185 (14%) |
Các miền gần với benchmark đứng đầu (một bản ghi có thể khớp nhiều miền): ngân hàng, đặt vé, bán lẻ, tìm kiếm web, toán, nhắn tin, du lịch, xã hội, bộ nhớ, giao dịch.
Chồng lấn với các benchmark
Quy tắc được áp dụng ở đây là quy tắc đã được yêu cầu: cùng miền với benchmark là được phép, lặp lại một mục kiểm thử thì không.
scripts/testset_index.py lập chỉ mục mọi prompt kiểm thử trong test_data/ — các
hướng dẫn *_test.jsonl của CAR-bench (các split *_train không phải mục kiểm thử
và không được lập chỉ mục), mọi hướng dẫn user-scenario của tau^3 và mọi lượt người
dùng trong dữ liệu BFCL V4 — dưới dạng chuỗi đã chuẩn hóa hoàn chỉnh và các cửa sổ 12
từ. Một ứng viên bị loại khi một lượt người dùng lặp lại toàn bộ prompt kiểm thử hoặc
chia sẻ từ hai cửa sổ 12 từ trở lên với prompt đó. Quy tắc này đã loại 418 bản ghi,
phần lớn đến từ nemotron_tool_calling (240) và toolmind_xlam (167).
Những gì quy tắc này cố ý không loại bỏ, và cần được nêu rõ khi báo cáo kết quả:
toolmind_tau_trainlà split retail train của tau-bench: cùng môi trường, công cụ và chính sách với tau^3 retail nhưng khác task. Có 5.000 bản ghi của split này trong biến thểno_reasoning.toolmind_apigen_mtchứa các quỹ đạo APIGen-MT được tạo trong môi trường retail và airline theo phong cách tau-bench.nemotron_customer_servicegồm 838 miền dịch vụ khách hàng tổng hợp theo dạng tau-bench (chính sách trong system prompt, xác thực → tra cứu → chỉnh sửa).
Vì vậy tau^3 retail và airline là cùng miền với dữ liệu này theo thiết kế. CAR-bench, BFCL Memory và truy xuất kiến thức ngân hàng của tau^3 vẫn thực sự chưa xuất hiện.
Cách tạo dữ liệu
# scan (song song, ~3 phút trên 25 GB) -> select -> build (~10 phút, tokenize mọi bản ghi)
python scripts/prepare_final_train_data.py all --workers 10
# kiểm tra độc lập: validator của fc_baselines cùng phép đo token mới
python scripts/verify_final_train_data.py
Build import trực tiếp fc_baselines.data.convert và fc_baselines.data.validate,
nên một bản ghi chỉ xuất hiện trong các file này khi cổng kiểm tra dữ liệu huấn luyện
chấp nhận nó. Kiểm tra độc lập chạy lại validator và tokenize lại mọi bản ghi: 0 bản
ghi bị từ chối, 0 id trùng, 0 bản ghi vượt cutoff_len, 0 khác biệt với độ dài token
được ghi trong meta. Kiểm tra này báo cáo 2.849 cảnh báo mềm
argument_type_mismatch trong no_reasoning/train.jsonl — các giá trị đối số có kiểu
không khớp schema của công cụ; theo hợp đồng, đây là cảnh báo chứ không phải lỗi, và
chúng đến từ các nguồn có khai báo kiểu lỏng. Các nguồn không thống nhất với định
dạng theo những cách cơ học được sửa thay vì loại bỏ (ToolMind không có call id và
bọc một số định nghĩa công cụ hai lần, Nemotron lưu arguments dưới dạng chuỗi
JSON); những thứ có thể thay đổi điều mô hình học được thì không được sửa — hội
thoại như vậy bị cắt về tiền tố hợp lệ dài nhất. Xem scripts/normalize_records.py.
Trong 1.360.511 bản ghi nguồn, 1.293.114 bản ghi sống sót sau chuẩn hóa. Các nguyên
nhân loại bỏ chính là 48.072 bản ghi dài hơn cutoff_len, 11.178 bản ghi không thể
làm cho hợp lệ bằng bất kỳ sửa chữa nào và 7.154 bản ghi gọi một công cụ không có
trong danh sách tools của chính chúng.
Hạn chế đã biết
- 477M token mỗi epoch trong
no_reasoning. Đây là một lần chạy SFT lớn; nếu cần ngân sách nhỏ hơn, hãy lấy mẫu con theometa.score(giảm dần). - Tìm kiếm web còn mỏng: chỉ 655 trong số 5.968 quỹ đạo tìm kiếm của Nemotron vừa với giới hạn 16.384 token. BFCL V4 Web Search sẽ được hỗ trợ yếu, còn BFCL Memory hoàn toàn không được bao phủ — không có nguồn nào ở đây có công cụ bộ nhớ lâu dài.
- Biến thể
reasoningthiên về Nemotron hơn. ToolMind chỉ gắn suy luận vào lượt assistant cuối cùng, nhưngreasoning_mode: reasoningyêu cầu suy luận ở mọi lượt gọi công cụ, nên phần lớn bản ghi nhiều lượt của ToolMind không hợp lệ ở đó (toolmind_tau_train: 48 trên 10.618). Phần thiếu được bù từnemotron_tool_calling, vì vậy nguồn này tăng lên 47.066 và biến thể reasoning có ít hội thoại dài hơn. - Mục tiêu 100.000 bản ghi cho ra 99.294 / 99.202 vì phép đếm token chính xác, được đo bằng chat template thực sau khi chọn, đã loại lần lượt 706 và 798 bản ghi.
- Downloads last month
- -