Datasets:
tools stringlengths 145 37.8k | messages listlengths 2 60 |
|---|---|
[{"type": "function", "function": {"strict": true, "name": "ping", "description": "Simple ping tool to test server responsiveness and prevent timeouts.", "parameters": {"properties": {}, "title": "pingArguments", "type": "object"}}}, {"type": "function", "function": {"strict": true, "name": "health_check", "description... | [
{
"role": "user",
"content": "What is the current weather in Chicago right now?",
"reasoning": null,
"tool_calls": null,
"tool_call_id": null
},
{
"role": "assistant",
"content": "",
"reasoning": "We need current weather in Chicago. Use get_current_weather tool; requires US locat... |
[{"type": "function", "function": {"strict": true, "name": "add", "description": "Adds two numbers together", "parameters": {"type": "object", "properties": {"firstNumber": {"type": "number", "description": "The first addend"}, "secondNumber": {"type": "number", "description": "The second addend"}}, "required": ["first... | [
{
"role": "user",
"content": "How can I calculate the total sales from the list of numbers [50, 75, 200, 120]?",
"reasoning": null,
"tool_calls": null,
"tool_call_id": null
},
{
"role": "assistant",
"content": "",
"reasoning": "The user asks: \"How can I calculate the total sales... |
[{"type": "function", "function": {"strict": true, "name": "historical_currency_data", "description": "Using the exchange rate on a certain historical date, converts from one currency to another.", "parameters": {"type": "object", "properties": {"from_currency": {"type": "string", "description": "Base currency code (AU... | [
{
"role": "user",
"content": "I'm considering a cultural immersion trip to Jamaica next year and I'm starting to research budgeting. Could you help me figure out what 5,000 USD would be in Jamaican dollars right now?",
"reasoning": null,
"tool_calls": null,
"tool_call_id": null
},
{
"rol... |
[{"type": "function", "function": {"strict": true, "name": "paper_rank", "description": "Using the method of paper rank to analyze the url", "parameters": {"type": "object", "properties": {"domains": {"type": "string", "description": "The domains you want to analyze."}, "api_key": {"type": "string", "description": "The... | [
{
"role": "user",
"content": "I was just putting together the monthly newsletter for our senior center's volunteer team and realized I need to share several website links about upcoming community events. I always get complaints that the links are too long and confusing in our printed bulletins. Do you know ... |
[{"type": "function", "function": {"strict": true, "name": "configure_api_key", "description": "Configure your Coinranking API key for this session. Get your key at https://account.coinranking.com/dashboard", "parameters": {"properties": {"api_key": {"description": "Your Coinranking API key", "type": "string"}}, "requi... | [
{
"role": "user",
"content": "I'm thinking about dollar-cost averaging into Bitcoin, planning to invest $500 every month. Can you help me determine how much Bitcoin I would own after 6 months if the price stays consistent at today's rate versus if it drops 10% each month? I want to compare the total amount ... |
[{"type": "function", "function": {"strict": true, "name": "drawing_generateCanvas", "description": "Generate a new drawing canvas with specified width and height.", "parameters": {"type": "object", "properties": {"width": {"type": "number", "description": "Width of the canvas in pixels"}, "height": {"type": "number", ... | [
{
"role": "user",
"content": "I need to create a visual element for my AI assistant that distinguishes emphasized areas in a report. Can you help me draw a red rectangle at coordinates (20, 30) with a size of 100x50 on my drawing canvas?",
"reasoning": null,
"tool_calls": null,
"tool_call_id": n... |
[{"type": "function", "function": {"strict": true, "name": "deleteaccount", "description": "Delete account API that takes a token parameter which is the authentication token of the user making the deletion request.", "parameters": {"type": "object", "properties": {"token": {"type": "string", "description": "The authent... | [
{
"role": "user",
"content": "Oh, I’m trying to get signed up for that new community arts portal — the one that lets us senior artists show our watercolors in virtual galleries. Could you help me register my account? My name is Jana Pentz, and I’ll be using my email jana.pentz74@gmail.com. Let’s set up a pa... |
[{"type": "function", "function": {"strict": true, "name": "sendim", "description": "Send an IM API that takes two parameters - 'recipient' and 'content'. The 'recipient' parameter refers to the recipient of the message and the 'content' parameter specifies the content of the message.", "parameters": {"type": "object",... | [
{
"role": "user",
"content": "Oh, I'm so glad you're here. I've been trying to get into this online art community platform—it's called \"Brush & Thread Hub\"—so I can post about our local Fontana Art Walk and start setting up that intergenerational quilting circle I've been dreaming about. But I must have f... |
[{"type": "function", "function": {"strict": true, "name": "get_technical_indicator_description", "description": "Get description of a technical indicator by an indicator symbol.", "parameters": {"type": "object", "properties": {"indicator_symbol": {"type": "string", "description": "The symbol of the technical indicato... | [
{
"role": "user",
"content": "We’re planning a community art fundraising project and need to analyze investment trends to propose a sustainable portfolio—can you describe what technical indicators I should consider when evaluating historic stock performance?",
"reasoning": null,
"tool_calls": null,
... |
[{"type": "function", "function": {"strict": true, "name": "clear_thoughts", "description": "Clear all recorded thoughts from the current session. Use this to start fresh if the thinking process needs to be reset.", "parameters": {"type": "object", "properties": {}}}}, {"type": "function", "function": {"strict": true, ... | [
{
"role": "user",
"content": "Could you give me a brief rundown of every decision and rationale we logged during the sprint planning meeting? I need a fast overview of the reasoning we captured.",
"reasoning": null,
"tool_calls": null,
"tool_call_id": null
},
{
"role": "assistant",
"... |
[{"type": "function", "function": {"strict": true, "name": "create_palace", "description": "Create a new memory palace with a specific theme", "parameters": {"type": "object", "properties": {"name": {"type": "string", "description": "The name of the memory palace"}, "description": {"type": "string", "description": "Des... | [
{
"role": "user",
"content": "I'm studying for my board exams and want to memorize the drug \"lisinopril\" using my memory palace techniques. I've already set up a \"Pharmacology Palace\" with a room named \"ACE Inhibitors\" for drugs that belong to that class. Could you retrieve the complete pharmacologic ... |
End of preview. Expand in Data Studio
Qwen3.8-Flash-Next Tool-Calling Dataset
A function/tool-calling preference and SFT dataset built from Nemotron-RL agentic prompt corpora. For each source prompt the model was sampled four times with thinking mode enabled; each candidate was scored against the dataset's ground-truth action by exact match on function name and parsed arguments.
Overview
| Stage | Count |
|---|---|
| Source prompts, Nemotron FC + Conv | 106,588 |
| Successfully sampled trajectories, n=4 | 106,574 |
| Total candidate completions | 426,296 |
| Auto-scored function-call ground-truth rows | 70,477 |
| Rows with at least one chosen candidate | 29,303 |
| Rows with a preference pair | 18,393 |
| Final SFT rows | 34,170 |
Source Datasets
| Tag | Dataset | Source rows | Rows kept |
|---|---|---|---|
| FC | nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1 |
9,620 | 9,606 |
| Conv | nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1 |
96,968 | 96,968 |
Sampling Protocol
Four samples per prompt with thinking enabled (reasoning is present on the scored candidates).
Difficulty Signal
Over the 70,477 auto-scored function-call rows:
- 0/4 — 58.4% (41,174 rows): no candidate matched the ground truth.
- 1–3/4 — 26.1% (18,393 rows): the preference-pair pool.
- 4/4 — 15.5% (10,910 rows): one candidate kept.
By source: FC 0/4 = 47.6% (2,340 / 4,918), Conv 0/4 = 59.2% (38,834 / 65,559).
| Chosen / 4 | Rows | Share |
|---|---|---|
| 0 | 41,174 | 58.4% |
| 1 | 7,475 | 10.6% |
| 2 | 5,465 | 7.8% |
| 3 | 5,453 | 7.7% |
| 4 | 10,910 | 15.5% |
At least one chosen candidate (pass@4): 41.6%. Median reasoning length is 141 (chosen) vs 279 (rejected) ≈tokens.
Failure breakdown on Conv function-call rows (262,236 candidates): chosen 26.5%, no tool call 31.2%, right function / wrong arguments 22.9%, wrong function 17.3%, multiple calls 2.1%.
Reasoning Length: Chosen vs Rejected
Across 281,908 auto-scored candidates (lengths are chars/4, a token estimate):
| Metric | Chosen | Rejected |
|---|---|---|
| Median | 141 | 279 |
| Mean | 227 | 397 |
Rejected reasoning is 75% longer on average. This is a marginal statistic and mostly reflects prompt difficulty, not a per-candidate quality signal. Controlling for the prompt (18,393 rows with both a chosen and a rejected candidate; for each row, the fraction of chosen-vs-rejected pairs in which the chosen reasoning is shorter, ties as ½), chosen is shorter in only 59.1% of pairs. Do not read the marginal gap as a recipe for length-based filtering.
finish_reason == "length" occurs in 160 of the 281,908 scored candidates (0.057%), all of them rejected.
740 candidates in total (0.17%) have neither content nor tool calls.
SFT Split Complexity
- Median tool inventory per row: 17 (max 53).
- Median trajectory length incl. system and tool turns: 7, p99 30.
SFT Files
| File | Rows | Description |
|---|---|---|
data/train.parquet |
34,170 | all selected rows |
Columns: ["tools", "messages"]; only the final assistant turn is the loss target.
Selection Rule for Function-Call Rows
| Chosen / Rejected / Unscored | Rule | Rows | SFT rows (before QC) |
|---|---|---|---|
| 4 / 0 / 0 | one random chosen | 10,910 | 10,910 |
| 3 / 1 / 0 | one random chosen | 5,453 | 5,453 |
| 2 / 2 / 0 | both chosen | 5,465 | 10,930 |
| 1 / 3 / 0 | the single chosen | 7,475 | 7,475 |
| 0 / 4 / 0 | drop | 41,174 | 0 |
| 0 / 0 / 4 | drop (message-only) | 36,097 | 0 |
QC removes 598 further candidates (72 refusals, 526 degenerate reasoning): 34,768 → 34,170.
Known Limitations
- Message-type ground truth is not auto-scored (36,097 rows); excluded from SFT.
- 0/4 rows are excluded, and rejected candidates are not included in this release.
- No LLM-judge quality pass. Exact-match scoring proves the action matches the ground truth, not that the reasoning is sound.
License and Attribution
Source prompts: NVIDIA Open Data License v1.0 (see the Nemotron-RL repos). Completions generated by Qwen3.8-Flash-Next — follow its model-card terms. Layout, scoring and chosen/rejected annotations: CC-BY-4.0.
- Downloads last month
- 119